Implement AI Host functionality with model caching and job management

- Added ai_host.py to handle AI tasks with FastAPI integration.
- Introduced HostState and ModelCache classes for managing client connections and model loading.
- Implemented job submission and cancellation endpoints in the AI host.
- Enhanced tests for AI host functionality, including client tracking and job cancellation.
- Updated existing tests to accommodate changes in the operator and command structure.
- Refactored runtime commands to use request-based architecture instead of command-line arguments.
- Improved model loading logic to handle CUDA memory management and reuse models efficiently.

README.md

@@ -9,15 +9,18 @@
 ## 当前功能
 
 - 使用 Blender Extension manifest，可通过 `Install from Disk` 安装
-- 在独立 Python Runtime 中运行 Torch 与 DA3，避免污染 Blender Python 进程
+- 在独立 Python Runtime 的常驻 HTTP AI Host 中运行 Torch、DA3 与 BEN2，避免污染 Blender Python 进程
+- Blender 使用内置 `urllib.request` 访问仅监听 `127.0.0.1` 的 Host；请求带随机 Token
+- 常驻 Host 复用 Torch、CUDA、BEN2 与最近使用的 DA3 模型，连续执行时不再重复加载
+- 多个 Blender 实例自动连接同一个共享 Host；`Host` 面板显示状态、端口、客户端与模型缓存，并提供安装、启动、清理、重启和停止控制
 - Runtime、模型和任务结果默认保存在 `~/.snapform`，可由多个 Blender 版本共享
 - 支持 DA3 Small、Base、Large 1.1、Giant 1.1 和 Nested Giant Large 1.1
 - 支持 BEN2 Base 单图背景剔除测试，输出透明 PNG 与推理耗时
 - 支持单张 JPG、JPEG、PNG、WEBP 图片或常见视频文件输入
 - 视频均匀抽帧，可设置参与重建的最大帧数
-- `Setup` 集中管理 Runtime，并逐项列出和下载所有可用模型
+- `Host > Models` 默认折叠，展开后可逐项查看和下载所有可用模型
 - Runtime 已就绪时不额外占用面板空间；模型列表每个模型只占一行
-- `Debug` 是独立且默认折叠的 AI 功能测试面板
+- `Debug` 默认折叠，内部的 DA3 与 BEN2 测试项也分别默认折叠
 - 使用一个导出类型下拉菜单和一个 `Export` 按钮
 - 只显示当前格式的专属参数；GLB 与 PLY 设置不会同时出现
 - Debug 的模型选项按格式过滤；PLY 只提供两个带 Gaussian head 的 Giant 模型
@@ -31,7 +34,7 @@
 - 转换对话框可选择推理模式、处理分辨率和面片网格细分
 - Debug 的 DA3 导出或 BEN2 背景剔除完成后只打开对应 job 文件夹，不向当前 Blender 文件添加任何数据
 - 完整保留每次任务生成的所有官方文件
-- 安装、模型下载和推理均在后台子进程运行，状态栏实时显示阶段、百分比进度条与取消按钮
+- 安装和模型下载使用后台子进程；推理提交给常驻 AI Host，状态栏实时显示阶段、百分比进度条与取消按钮
 - 后台任务运行时，Install、Download 和 Export 按钮保持原位置并变灰，任务完成后恢复
 
 ## 使用
@@ -39,19 +42,20 @@
 1. 使用 `uv run python pack.py` 构建 Extension ZIP。
 2. 在 Blender 的 `Preferences > Extensions` 中选择 `Install from Disk`。
 3. 在 Extension Preferences 中确认 `uv` 路径和 Torch Backend。
-4. 点击 `Install Runtime`。首次安装会下载 Python、PyTorch、DA3 和 BEN2 依赖。
+4. 在 `Host` 点击 `Install Runtime`。首次安装会下载 Python、PyTorch、DA3、BEN2、FastAPI、Uvicorn，
+   并默认下载 `DA3 Small`、`DA3 Base` 和 `BEN2 Base` 模型。
 5. 在 3D View 按 `N`，打开 `SnapForm` 页签。
-6. 在 `Setup` 中找到需要的模型并点击该行的 `Download`。
+6. 如需其他模型，展开 `Host > Models`，找到模型并点击该行的 `Download`。
 7. 展开默认收起的 `Debug`，选择 `Export Format` 和本次执行使用的模型。
 8. 选择一张图片或一个视频文件。视频可设置均匀抽取的帧数，然后点击 `Export`。
-9. 测试背景剔除时，在 `Setup` 下载 `BEN2 Base`，然后在 `Debug` 的
+9. 测试背景剔除时，确认 `Host > Models` 中的 `BEN2 Base` 已就绪，然后在 `Debug` 的
    `BEN2 Background Removal` 区域选择单张图片并点击 `Remove Background`。
    完成后会打开 job 文件夹，其中的 `foreground.png` 带透明 Alpha；状态栏显示
    BEN2 推理、模型加载与任务总耗时。当前测试固定关闭较慢的 foreground refinement。
 
 ## Image Empty 去背景
 
-1. 在 `Setup` 下载 `BEN2 Base`。
+1. 确认 `Host > Models` 中的 `BEN2 Base` 已就绪；首次安装 Runtime 时会默认下载。
 2. 在 3D View 选中 Image Empty，右键选择 `Remove Background`。
 3. 完成后以 Fake User 保留原 Image datablock，新建并 Pack 一张透明 PNG 到当前 `.blend`，然后只把
    当前 Empty 的 `data` 换成新图；Empty 的 Transform、显示尺寸、Offset 及其他设置保持不变。
@@ -66,16 +70,17 @@
 1. 在 3D View 选中 `Object Type = Empty`、`Display As = Image` 的对象。
 2. 右键选择 `Convert to Depth Plane`，通过 `Depth from` 选择深度来源：
    - `AI Generated`：显示 Model、Use Ray Pose 和 Resolution Scale。实际 Process Res 为图片长边像素数
-     乘以 Resolution Scale；范围为 `0.1–1.0`，默认 `0.5`。计算后的 Process Res 会直接显示在
+     乘以 Resolution Scale；范围为 `0.1–1.0`，默认 `1.0`。计算后的 Process Res 会直接显示在
      `Resolution Scale (… px)` 标签中。
      转换导出 `mini_npz`，使用前需要在
-     `Setup` 下载所选模型，并固定使用 `process_res_method="upper_bound_resize"`；设备和参考视图策略
+     `Host > Models` 下载所选模型，并固定使用 `process_res_method="upper_bound_resize"`；设备和参考视图策略
-     继续使用场景设置。
+     继续使用场景设置。Model 下拉菜单只列出已经下载的 DA3 模型；若没有可用模型，对话框会引导到
+     `Host > Models` 下载。
    - `Original`：直接把原图作为深度贴图，不显示估算参数。Remove Background 关闭时不启动
      Runtime，也不要求源图片存在于磁盘。
    `Original` 是默认深度来源。选择 `AI Generated` 时，其 Model、Use Ray Pose 与 Resolution Scale
-   参数会集中显示在独立边框中。两种模式始终显示 Remove Background、Displace via、Unshaded、
+   参数会集中显示在独立边框中。两种模式都有默认折叠的 Options，内部依次为 Displace via、
-   Subdivision 与 Delete Original。
+   Subdivision、Remove Background 和 Unshaded；Delete Original 保持在折叠项外。
    Remove Background 默认关闭，与 Depth from 无关；开启后需要下载 `BEN2 Base`，并从原图生成、
    Pack `.rb` 透明图片。它只替换生成面片材质的颜色图片，不修改源 Image Empty 的 Image datablock。
    AI Generated 仍从原图生成 `.depth`；例如 `img.png` 会得到 `img.rb` 与 `img.depth`，不会生成
@@ -130,8 +135,9 @@ Principled BSDF 的 `Base Color`，并通过
 `0.001`，会乘以面片 Transform 的统一缩放倍率。原图 Color 还会经过 `Float Curve` 连接到
 Principled BSDF 的 `Roughness`，曲线两端输出分别为 `0.5` 和 `0.1`；材质 IOR 保持
 Principled BSDF 默认值 `1.5`。
-开启 `Unshaded` 后，Surface 改为原图 Color 直接连接 Emission，再连接 Material Output；不创建
+开启 `Unshaded` 后，原图 Color 连接 Emission，原图 Alpha 作为 Mix Shader 系数，在
-Principled BSDF、Bump 或 Float Curve。若 Displace via 选择 `Material`，深度贴图与 Displacement
+Transparent BSDF 与 Emission 之间混合后连接 Material Output；不创建 Principled BSDF、Bump 或
+Float Curve。若 Displace via 选择 `Material`，深度贴图与 Displacement
 节点仍会保留并连接 Material Output 的 Displacement。
 
 这里的“原始距离”是 DA3 模型输出的尺度。普通 Small、Base、Large 和 Giant 模型的深度不应
@@ -219,6 +225,7 @@ Extension 本身不包含 Torch。`Install Runtime` 使用 `uv` 创建独立 Pyt
 - xFormers 0.0.33.post2（与 PyTorch 2.9.1 对应）
 - 固定到已验证 commit 的 Depth Anything 3
 - 固定到已验证 commit 的 BEN2
+- FastAPI 0.139.2 与 Uvicorn 0.51.0
 - 固定到 DA3 官方推荐 commit 的 `gsplat`；仅安装 PLY 导出所需的纯 Python 部分，
   不编译本机 CUDA rasterizer
 - `addict` 2.4.0（补充 DA3 上游包元数据暂未声明的运行依赖）
@@ -226,19 +233,43 @@ Extension 本身不包含 Torch。`Install Runtime` 使用 `uv` 创建独立 Pyt
 默认 CUDA Backend 为 CUDA 12.8，也可以选择 CPU。CPU 可以用于验证，但实际推理
 通常很慢。
 
+Extension 启用后只会查找已有的共享 AI Host，不会主动启动新进程。第一次执行 AI 操作或在侧边栏
+手动点击 `Start` 时才会启动 Host。Host 只绑定回环地址，并让操作系统分配动态端口；第一个实例把
+端口、进程和随机 Token 写入 `host.json`，其余 Blender 通过标准库
+`urllib.request` 连接同一个 Host。`host.lock` 保证同一 Storage Root 同时只有一个 Host，Token
+则阻止同机其他进程直接调用接口。Host 全局同时只执行一个 AI 任务。
+
+每个 Blender 会注册独立客户端并每 5 秒发送心跳。正常关闭 Blender 时仅注销自身，不会直接终止
+其他 Blender 正在使用的 Host；异常退出的客户端会在 20 秒后过期。最后一个客户端离开后，Host
+保留模型缓存等待 60 秒，再在空闲时自动退出。取消任务不会重启 Host，并会保留已经加载的模型；
+如果取消发生在一次不可中断的 Torch 调用中，任务会在该调用返回后的最近检查点结束。
+
+侧边栏 `Host` 面板以 `Running (Clients: n)` 表示空闲服务，以 `Running (Busy)` 表示正在执行任务，
+并显示缓存模型和活动任务。空闲且只有当前 Blender 连接时，可以清理模型缓存、重启或停止 Host；
+多客户端或任务执行期间不会允许破坏共享服务。
+
+BEN2 在首次使用后保持加载；DA3 保留最近使用的一个模型。切换 DA3 模型时会释放旧模型；加载时如果
+显存不足，会先释放另一类缓存模型并重试。
+
+Runtime Schema 已升级，旧 Runtime 需要重新执行 `Install Runtime` 才能安装 HTTP Host 依赖。
+Host 的启动和崩溃日志写入 `~/.snapform/ai-host.log`。
+
 Runtime、模型和任务结果默认保存在用户主目录的 `~/.snapform`：
 
 ```text
 ~/.snapform/
 ├── runtime/
 ├── models/
-└── jobs/
+├── jobs/
+├── ai-host.log
+├── host.json       # Host 运行时存在
+└── host.lock       # Host 运行时持有
 ```
 
 不同 Blender 版本安装相同 Extension 后会共用这个目录。也可以在 Extension
 Preferences 的 `Storage Root` 中修改共享位置。不要把模型写入 Extension 安装目录，
 因为该目录可能只读，并且升级时会被替换。不要同时从多个 Blender 实例执行
-`Install Runtime`；推理任务和模型读取可以共享。
+`Install Runtime`；推理 Host、任务队列和模型读取可以共享。
 
 ## 开发
 
@@ -255,6 +286,7 @@ uv run python pack.py
 ├── src/
 │   └── snapform/
 │       ├── blender_manifest.toml
+│       ├── ai_host.py
 │       ├── bootstrap_runtime.py
 │       ├── models.py
 │       ├── image_depth.py

src/snapform/__init__.py

@@ -1,5 +1,6 @@
 import bpy
 
+from . import runtime
 from .image_depth import (
     ApplyRemovedBackground,
     BuildDepthPlane,
@@ -10,11 +11,18 @@ from .image_depth import (
 )
 from .operators import (
     CancelJob,
+    ControlHost,
     RunJob,
     cancel_active_job,
     close_active_job,
 )
-from .panel import DebugPanel, RuntimeModelsPanel
+from .panel import (
+    Ben2DebugPanel,
+    Da3DebugPanel,
+    DebugPanel,
+    HostPanel,
+    ModelsPanel,
+)
 from .preferences import SnapFormPreferences
 from .properties import SnapFormSettings
 from .status_bar import draw_status_bar
@@ -25,13 +33,17 @@ CLASSES = (
     SnapFormSettings,
     RunJob,
     CancelJob,
+    ControlHost,
     ApplyRemovedBackground,
     RemoveImageBackground,
     BuildDepthPlane,
     BuildOriginalDepthPlane,
     ConvertImageToDepthPlane,
-    RuntimeModelsPanel,
+    HostPanel,
+    ModelsPanel,
     DebugPanel,
+    Da3DebugPanel,
+    Ben2DebugPanel,
 )
 
 
@@ -43,11 +55,14 @@ def register():
     )
     bpy.types.STATUSBAR_HT_header.append(draw_status_bar)
     bpy.types.VIEW3D_MT_object_context_menu.append(draw_image_context_menu)
+    runtime.schedule_host_start()
 
 
 def unregister():
     cancel_active_job()
     close_active_job()
+    runtime.unschedule_host()
+    runtime.detach_host_client()
     bpy.types.VIEW3D_MT_object_context_menu.remove(draw_image_context_menu)
     bpy.types.STATUSBAR_HT_header.remove(draw_status_bar)
     if hasattr(bpy.types.Scene, "snapform_settings"):

src/snapform/ai_host.py

@@ -0,0 +1,572 @@
+import argparse
+import gc
+import hmac
+import json
+import os
+import sys
+import threading
+import time
+import traceback
+from contextlib import asynccontextmanager
+from pathlib import Path
+from types import SimpleNamespace
+
+
+SCRIPT_DIRECTORY = Path(__file__).resolve().parent
+if str(SCRIPT_DIRECTORY) not in sys.path:
+    sys.path.insert(0, str(SCRIPT_DIRECTORY))
+
+import worker
+
+
+HOST_SCHEMA = 2
+CLIENT_TIMEOUT_SECONDS = 20.0
+IDLE_SHUTDOWN_SECONDS = 60.0
+
+
+class HostLock:
+    def __init__(self, path):
+        self.path = Path(path)
+        self.file = None
+
+    def acquire(self):
+        self.path.parent.mkdir(parents=True, exist_ok=True)
+        lock_file = self.path.open("a+b")
+        lock_file.seek(0, os.SEEK_END)
+        if lock_file.tell() == 0:
+            lock_file.write(b"0")
+            lock_file.flush()
+        lock_file.seek(0)
+        try:
+            if sys.platform == "win32":
+                import msvcrt
+
+                msvcrt.locking(lock_file.fileno(), msvcrt.LK_NBLCK, 1)
+            else:
+                import fcntl
+
+                fcntl.flock(lock_file.fileno(), fcntl.LOCK_EX | fcntl.LOCK_NB)
+        except OSError:
+            lock_file.close()
+            return False
+        self.file = lock_file
+        return True
+
+    def release(self):
+        if self.file is None:
+            return
+        try:
+            self.file.seek(0)
+            if sys.platform == "win32":
+                import msvcrt
+
+                msvcrt.locking(self.file.fileno(), msvcrt.LK_UNLCK, 1)
+            else:
+                import fcntl
+
+                fcntl.flock(self.file.fileno(), fcntl.LOCK_UN)
+        finally:
+            self.file.close()
+            self.file = None
+
+
+class ModelCache:
+    def __init__(self):
+        import torch
+
+        self.torch = torch
+        self.ben2_model = None
+        self.ben2_key = None
+        self.da3_model = None
+        self.da3_key = None
+
+    def get_ben2(self, model_dir, device):
+        key = (str(Path(model_dir).resolve()), device.type)
+        if self.ben2_model is not None and self.ben2_key == key:
+            return self.ben2_model, 0.0
+        self.release_ben2()
+        try:
+            return self._load_ben2(model_dir, device, key)
+        except self.torch.cuda.OutOfMemoryError:
+            self.release_da3()
+            return self._load_ben2(model_dir, device, key)
+
+    def _load_ben2(self, model_dir, device, key):
+        from ben2 import BEN_Base
+
+        started = time.perf_counter()
+        model = BEN_Base.from_pretrained(model_dir)
+        model = model.to(device=device)
+        model.eval()
+        self.ben2_model = model
+        self.ben2_key = key
+        elapsed_ms = (time.perf_counter() - started) * 1000.0
+        return model, elapsed_ms
+
+    def get_da3(self, model_dir, device):
+        key = (str(Path(model_dir).resolve()), device.type)
+        if self.da3_model is not None and self.da3_key == key:
+            return self.da3_model
+        self.release_da3()
+        try:
+            return self._load_da3(model_dir, device, key)
+        except self.torch.cuda.OutOfMemoryError:
+            self.release_ben2()
+            return self._load_da3(model_dir, device, key)
+
+    def _load_da3(self, model_dir, device, key):
+        from depth_anything_3.api import DepthAnything3
+
+        model = DepthAnything3.from_pretrained(model_dir)
+        model = model.to(device=device)
+        model.eval()
+        self.da3_model = model
+        self.da3_key = key
+        return model
+
+    def release_ben2(self):
+        self.ben2_model = None
+        self.ben2_key = None
+        self._release_cuda_cache()
+
+    def release_da3(self):
+        self.da3_model = None
+        self.da3_key = None
+        self._release_cuda_cache()
+
+    def close(self):
+        self.ben2_model = None
+        self.da3_model = None
+        self.ben2_key = None
+        self.da3_key = None
+        self._release_cuda_cache()
+
+    def status(self):
+        return {
+            "ben2": (
+                Path(self.ben2_key[0]).name if self.ben2_key is not None else ""
+            ),
+            "da3": (
+                f"da3_{Path(self.da3_key[0]).name}"
+                if self.da3_key is not None
+                else ""
+            ),
+        }
+
+    def _release_cuda_cache(self):
+        gc.collect()
+        if self.torch.cuda.is_available():
+            self.torch.cuda.empty_cache()
+
+
+class HostState:
+    def __init__(self, token, instance_id):
+        self.token = token
+        self.instance_id = instance_id
+        self.models = ModelCache()
+        self.lock = threading.Lock()
+        self.clients = {}
+        self.active_job = None
+        self.started_at = time.time()
+        self.last_without_clients = time.monotonic()
+        self.shutdown_event = threading.Event()
+        self.monitor_stop = threading.Event()
+        self.monitor_thread = None
+
+    def register_client(self, client_id, label):
+        with self.lock:
+            self.clients[client_id] = {
+                "label": label,
+                "last_seen": time.monotonic(),
+            }
+            self.last_without_clients = None
+
+    def heartbeat(self, client_id):
+        with self.lock:
+            client = self.clients.get(client_id)
+            if client is None:
+                return False
+            client["last_seen"] = time.monotonic()
+            return True
+
+    def unregister_client(self, client_id):
+        with self.lock:
+            self.clients.pop(client_id, None)
+            self._mark_empty_clients()
+
+    def reserve(self, job_id, client_id, command, status_path):
+        with self.lock:
+            if client_id not in self.clients or self.active_job is not None:
+                return False
+            self.active_job = {
+                "job_id": job_id,
+                "client_id": client_id,
+                "command": command,
+                "status": status_path,
+                "cancel_requested": False,
+            }
+            return True
+
+    def snapshot(self):
+        self.expire_clients()
+        with self.lock:
+            return {
+                "ready": True,
+                "host_schema": HOST_SCHEMA,
+                "instance_id": self.instance_id,
+                "busy": self.active_job is not None,
+                "active_job": dict(self.active_job) if self.active_job else None,
+                "clients": len(self.clients),
+                "cached_models": self.models.status(),
+                "started_at": self.started_at,
+            }
+
+    def clear_models(self):
+        with self.lock:
+            if self.active_job is not None:
+                return False
+            self.models.close()
+            return True
+
+    def request_shutdown(self, client_id):
+        with self.lock:
+            if client_id not in self.clients:
+                return "missing"
+            if len(self.clients) > 1:
+                return "shared"
+            if self.active_job is not None:
+                return "busy"
+            self.shutdown_event.set()
+            return "accepted"
+
+    def cancel_job(self, job_id, client_id):
+        with self.lock:
+            active = self.active_job
+            if active is None or active["job_id"] != job_id:
+                return None
+            if active["client_id"] != client_id:
+                return None
+            active["cancel_requested"] = True
+            return dict(active)
+
+    def is_job_cancelled(self, job_id):
+        with self.lock:
+            active = self.active_job
+            return bool(
+                active is not None
+                and active["job_id"] == job_id
+                and active["cancel_requested"]
+            )
+
+    def run_job(self, payload):
+        job_id = payload["job_id"]
+        parameters = payload["parameters"]
+        status_path = parameters["status"]
+        worker.write_status(
+            status_path,
+            0.0,
+            "AI Host accepted the task",
+            state="running",
+        )
+        try:
+            cancel_check = lambda: self.is_job_cancelled(job_id)
+            worker.check_cancelled(cancel_check)
+            args = SimpleNamespace(**parameters)
+            if payload["command"] == "remove-background":
+                worker.remove_background(
+                    args,
+                    model_cache=self.models,
+                    cancel_check=cancel_check,
+                )
+            else:
+                worker.reconstruct(
+                    args,
+                    model_cache=self.models,
+                    cancel_check=cancel_check,
+                )
+            worker.check_cancelled(cancel_check)
+        except worker.TaskCancelled:
+            worker.write_status(
+                status_path,
+                1.0,
+                "Task cancelled",
+                state="cancelled",
+            )
+        except Exception as error:
+            traceback.print_exc()
+            worker.write_status(
+                status_path,
+                1.0,
+                "Task failed",
+                str(error),
+                state="failed",
+            )
+        finally:
+            with self.lock:
+                if (
+                    self.active_job is not None
+                    and self.active_job["job_id"] == job_id
+                ):
+                    self.active_job = None
+                self._mark_empty_clients()
+
+    def start_monitor(self):
+        self.monitor_thread = threading.Thread(
+            target=self._monitor,
+            name="SnapFormHostMonitor",
+            daemon=True,
+        )
+        self.monitor_thread.start()
+
+    def close(self):
+        self.monitor_stop.set()
+        if self.monitor_thread is not None:
+            self.monitor_thread.join(timeout=1.0)
+        self.models.close()
+
+    def expire_clients(self):
+        deadline = time.monotonic() - CLIENT_TIMEOUT_SECONDS
+        with self.lock:
+            expired = [
+                client_id
+                for client_id, client in self.clients.items()
+                if client["last_seen"] < deadline
+            ]
+            for client_id in expired:
+                self.clients.pop(client_id, None)
+            self._mark_empty_clients()
+
+    def _mark_empty_clients(self):
+        if self.clients:
+            self.last_without_clients = None
+        elif self.last_without_clients is None:
+            self.last_without_clients = time.monotonic()
+
+    def _monitor(self):
+        while not self.monitor_stop.wait(1.0):
+            self.expire_clients()
+            with self.lock:
+                idle = self.active_job is None
+                empty_since = self.last_without_clients
+            if not idle or empty_since is None:
+                continue
+            if time.monotonic() - empty_since >= IDLE_SHUTDOWN_SECONDS:
+                self.shutdown_event.set()
+                return
+
+
+def create_app(token, instance_id="test"):
+    from fastapi import BackgroundTasks, FastAPI, Header, HTTPException
+
+    state = HostState(token, instance_id)
+
+    @asynccontextmanager
+    async def lifespan(_app):
+        state.start_monitor()
+        yield
+        state.close()
+
+    app = FastAPI(
+        title="SnapForm AI Host",
+        docs_url=None,
+        redoc_url=None,
+        openapi_url=None,
+        lifespan=lifespan,
+    )
+    app.state.snapform_host_state = state
+
+    def authorize(x_snapform_token: str = Header(default="")):
+        if not hmac.compare_digest(x_snapform_token, state.token):
+            raise HTTPException(status_code=401, detail="Invalid host token")
+
+    @app.get("/health")
+    def health(x_snapform_token: str = Header(default="")):
+        authorize(x_snapform_token)
+        return state.snapshot()
+
+    @app.post("/clients", status_code=201)
+    def register_client(
+        payload: dict,
+        x_snapform_token: str = Header(default=""),
+    ):
+        authorize(x_snapform_token)
+        client_id = str(payload.get("client_id", ""))
+        if not client_id:
+            raise HTTPException(status_code=422, detail="Missing client id")
+        state.register_client(client_id, str(payload.get("label", "Blender")))
+        return {"client_id": client_id, "lease_seconds": CLIENT_TIMEOUT_SECONDS}
+
+    @app.post("/clients/{client_id}/heartbeat")
+    def heartbeat(
+        client_id: str,
+        x_snapform_token: str = Header(default=""),
+    ):
+        authorize(x_snapform_token)
+        if not state.heartbeat(client_id):
+            raise HTTPException(status_code=404, detail="Client is not registered")
+        return {"client_id": client_id, "state": "active"}
+
+    @app.delete("/clients/{client_id}")
+    def unregister_client(
+        client_id: str,
+        x_snapform_token: str = Header(default=""),
+    ):
+        authorize(x_snapform_token)
+        state.unregister_client(client_id)
+        return {"client_id": client_id, "state": "removed"}
+
+    @app.post("/jobs", status_code=202)
+    def submit_job(
+        payload: dict,
+        background_tasks: BackgroundTasks,
+        x_snapform_token: str = Header(default=""),
+    ):
+        authorize(x_snapform_token)
+        job_id = str(payload.get("job_id", ""))
+        client_id = str(payload.get("client_id", ""))
+        command = payload.get("command")
+        parameters = payload.get("parameters")
+        if not job_id or command not in {"remove-background", "reconstruct"}:
+            raise HTTPException(status_code=422, detail="Invalid job request")
+        if not isinstance(parameters, dict) or not parameters.get("status"):
+            raise HTTPException(status_code=422, detail="Missing job parameters")
+        if not state.reserve(
+            job_id,
+            client_id,
+            command,
+            parameters["status"],
+        ):
+            raise HTTPException(status_code=409, detail="AI Host is busy")
+        worker.write_status(
+            parameters["status"],
+            0.0,
+            "Task queued",
+            state="queued",
+        )
+        background_tasks.add_task(state.run_job, payload)
+        return {"job_id": job_id, "state": "queued"}
+
+    @app.delete("/jobs/{job_id}", status_code=202)
+    def cancel_job(
+        job_id: str,
+        payload: dict,
+        x_snapform_token: str = Header(default=""),
+    ):
+        authorize(x_snapform_token)
+        client_id = str(payload.get("client_id", ""))
+        active = state.cancel_job(job_id, client_id)
+        if active is None:
+            raise HTTPException(status_code=404, detail="Active job was not found")
+        try:
+            current = json.loads(
+                Path(active["status"]).read_text(encoding="utf-8")
+            )
+            progress = float(current.get("progress", 0.0))
+        except (OSError, ValueError, TypeError, json.JSONDecodeError):
+            progress = 0.0
+        worker.write_status(
+            active["status"],
+            progress,
+            "Cancelling task",
+            state="cancelling",
+        )
+        return {"job_id": job_id, "state": "cancelling"}
+
+    @app.post("/models/clear")
+    def clear_models(x_snapform_token: str = Header(default="")):
+        authorize(x_snapform_token)
+        if not state.clear_models():
+            raise HTTPException(status_code=409, detail="AI Host is busy")
+        return {"state": "cleared"}
+
+    @app.post("/shutdown", status_code=202)
+    def shutdown_host(
+        payload: dict,
+        x_snapform_token: str = Header(default=""),
+    ):
+        authorize(x_snapform_token)
+        result = state.request_shutdown(str(payload.get("client_id", "")))
+        if result == "shared":
+            raise HTTPException(status_code=409, detail="Other Blender clients are connected")
+        if result == "busy":
+            raise HTTPException(status_code=409, detail="AI Host is busy")
+        if result != "accepted":
+            raise HTTPException(status_code=404, detail="Client is not registered")
+        return {"state": "stopping"}
+
+    return app
+
+
+def parse_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--host", default="127.0.0.1")
+    parser.add_argument("--port", type=int, required=True)
+    parser.add_argument("--token", required=True)
+    parser.add_argument("--instance-id", required=True)
+    parser.add_argument("--discovery", required=True)
+    parser.add_argument("--lock", required=True)
+    return parser.parse_args()
+
+
+def write_discovery(args):
+    path = Path(args.discovery)
+    path.parent.mkdir(parents=True, exist_ok=True)
+    temporary = path.with_name(f"{path.name}.tmp")
+    temporary.write_text(
+        json.dumps(
+            {
+                "host_schema": HOST_SCHEMA,
+                "port": args.port,
+                "token": args.token,
+                "pid": os.getpid(),
+                "instance_id": args.instance_id,
+            }
+        ),
+        encoding="utf-8",
+    )
+    temporary.replace(path)
+
+
+def remove_discovery(path, instance_id):
+    path = Path(path)
+    try:
+        payload = json.loads(path.read_text(encoding="utf-8"))
+    except (FileNotFoundError, json.JSONDecodeError, OSError):
+        return
+    if payload.get("instance_id") != instance_id:
+        return
+    path.unlink(missing_ok=True)
+
+
+def main():
+    import uvicorn
+
+    args = parse_args()
+    host_lock = HostLock(args.lock)
+    if not host_lock.acquire():
+        return
+    try:
+        app = create_app(args.token, args.instance_id)
+        config = uvicorn.Config(
+            app,
+            host=args.host,
+            port=args.port,
+            access_log=False,
+            log_level="warning",
+        )
+        server = uvicorn.Server(config)
+
+        def watch_shutdown():
+            app.state.snapform_host_state.shutdown_event.wait()
+            server.should_exit = True
+
+        threading.Thread(target=watch_shutdown, daemon=True).start()
+        write_discovery(args)
+        server.run()
+    finally:
+        remove_discovery(args.discovery, args.instance_id)
+        host_lock.release()
+
+
+if __name__ == "__main__":
+    main()

src/snapform/blender_manifest.toml

@@ -12,7 +12,7 @@ platforms = ["windows-x64"]
 license = ["SPDX:GPL-3.0-or-later"]
 
 [permissions]
-network = "Download DA3 runtime dependencies and model weights"
+network = "Access the local AI Host and download runtime dependencies and model weights"
 files = "Read source images and store runtime data"
 
 [build]

src/snapform/bootstrap_runtime.py

@@ -6,16 +6,26 @@ import sys
 from pathlib import Path
 
 
+SCRIPT_DIRECTORY = Path(__file__).resolve().parent
+if str(SCRIPT_DIRECTORY) not in sys.path:
+    sys.path.insert(0, str(SCRIPT_DIRECTORY))
+
+from models import BEN2_MODEL, MODELS
+
+
 TORCH_VERSION = "2.9.1"
 TORCHVISION_VERSION = "0.24.1"
 XFORMERS_VERSION = "0.0.33.post2"
 ADDICT_VERSION = "2.4.0"
+FASTAPI_VERSION = "0.139.2"
+UVICORN_VERSION = "0.51.0"
 GSPLAT_COMMIT = "0b4dddf04cb687367602c01196913cde6a743d70"
-RUNTIME_SCHEMA = 3
+RUNTIME_SCHEMA = 4
 TORCH_INDEXES = {
     "cu128": "https://download.pytorch.org/whl/cu128",
     "cpu": "https://download.pytorch.org/whl/cpu",
 }
+DEFAULT_MODEL_KEYS = ("SMALL", "BASE", BEN2_MODEL.key)
 
 
 def write_status(path, progress, message, error=None):
@@ -46,6 +56,33 @@ def environment_python(runtime_dir):
     return Path(runtime_dir) / "bin" / "python"
 
 
+def download_default_models(args, python):
+    models = {**MODELS, BEN2_MODEL.key: BEN2_MODEL}
+    download_script = (
+        "import json, sys; "
+        "from huggingface_hub import snapshot_download; "
+        "patterns = json.loads(sys.argv[3]); "
+        "snapshot_download(repo_id=sys.argv[1], local_dir=sys.argv[2], "
+        "allow_patterns=patterns or None)"
+    )
+    models_dir = Path(args.models_dir)
+    models_dir.mkdir(parents=True, exist_ok=True)
+    for index, model_key in enumerate(DEFAULT_MODEL_KEYS):
+        model = models[model_key]
+        progress = 0.90 + index * 0.03
+        write_status(args.status, progress, f"Downloading {model.label}")
+        run(
+            [
+                python,
+                "-c",
+                download_script,
+                model.repository,
+                models_dir / model.key.lower(),
+                json.dumps(model.download_patterns),
+            ]
+        )
+
+
 def install(args):
     runtime_dir = Path(args.runtime_dir)
     runtime_dir.parent.mkdir(parents=True, exist_ok=True)
@@ -129,6 +166,19 @@ def install(args):
         ]
     )
 
+    write_status(args.status, 0.75, "Installing AI Host")
+    run(
+        [
+            args.uv,
+            "pip",
+            "install",
+            "--python",
+            python,
+            f"fastapi=={FASTAPI_VERSION}",
+            f"uvicorn=={UVICORN_VERSION}",
+        ]
+    )
+
     write_status(args.status, 0.78, "Installing gsplat build tools")
     run(
         [
@@ -162,7 +212,7 @@ def install(args):
         gsplat_environment,
     )
 
-    write_status(args.status, 0.92, "Checking Runtime")
+    write_status(args.status, 0.88, "Checking Runtime")
     run(
         [
             python,
@@ -171,12 +221,15 @@ def install(args):
                 "import torch; "
                 "from depth_anything_3.api import DepthAnything3; "
                 "from ben2 import BEN_Base; "
+                "import fastapi, uvicorn; "
                 "import gsplat; "
                 "print(torch.__version__, torch.cuda.is_available())"
             ),
         ]
     )
 
+    download_default_models(args, python)
+
     ready_path = runtime_dir / "snapform_ready.json"
     ready_path.write_text(
         json.dumps(
@@ -185,6 +238,8 @@ def install(args):
                 "python": "3.12",
                 "torch": TORCH_VERSION,
                 "xformers": XFORMERS_VERSION,
+                "fastapi": FASTAPI_VERSION,
+                "uvicorn": UVICORN_VERSION,
                 "backend": args.backend,
                 "da3_commit": args.da3_commit,
                 "ben2_commit": args.ben2_commit,
@@ -202,6 +257,7 @@ def parse_args():
     parser = argparse.ArgumentParser()
     parser.add_argument("--uv", required=True)
     parser.add_argument("--runtime-dir", required=True)
+    parser.add_argument("--models-dir", required=True)
     parser.add_argument("--backend", choices=tuple(TORCH_INDEXES), required=True)
     parser.add_argument("--da3-commit", required=True)
     parser.add_argument("--ben2-commit", required=True)

src/snapform/image_depth.py

@@ -26,8 +26,13 @@ IMAGE_NAME_SUFFIXES = (
 
 def image_depth_model_items(owner, context):
     from .models import model_enum_items
+    from .runtime import model_ready
 
-    return model_enum_items(owner, context)
+    return [
+        item
+        for item in model_enum_items(owner, context)
+        if model_ready(item[0])
+    ]
 
 
 _last_depth_plane_model = "BASE"
@@ -921,8 +926,13 @@ class RemoveImageBackground(bpy.types.Operator):
     @classmethod
     def poll(cls, context):
         from .operators import active_job
+        from .runtime import host_busy
 
-        return is_image_empty(context.object) and active_job() is None
+        return (
+            is_image_empty(context.object)
+            and active_job() is None
+            and not host_busy()
+        )
 
     def execute(self, context):
         from .operators import IMAGE_BACKGROUND_JOB
@@ -1001,7 +1011,7 @@ class ConvertImageToDepthPlane(bpy.types.Operator):
     snapform_process_res_factor: bpy.props.FloatProperty(
         name="Resolution Scale",
         description="Scale applied to the source image long-edge resolution",
-        default=0.5,
+        default=1.0,
         min=0.1,
         max=1.0,
         step=10,
@@ -1019,18 +1029,30 @@ class ConvertImageToDepthPlane(bpy.types.Operator):
         min=0,
         max=10,
     )
-
     @classmethod
     def poll(cls, context):
         from .operators import active_job
+        from .runtime import host_busy
 
-        return is_image_empty(context.object) and active_job() is None
+        return (
+            is_image_empty(context.object)
+            and active_job() is None
+            and not host_busy()
+        )
 
     def invoke(self, context, _event):
         global _last_depth_plane_model
 
         if not self.properties.is_property_set("snapform_model"):
-            self.snapform_model = _last_depth_plane_model
+            available_keys = [
+                item[0] for item in image_depth_model_items(self, context)
+            ]
+            if available_keys:
+                self.snapform_model = (
+                    _last_depth_plane_model
+                    if _last_depth_plane_model in available_keys
+                    else available_keys[0]
+                )
         return context.window_manager.invoke_props_dialog(self, width=340)
 
     def draw(self, context):
@@ -1038,6 +1060,11 @@ class ConvertImageToDepthPlane(bpy.types.Operator):
         layout.prop(self, "snapform_depth_source")
         if self.snapform_depth_source == "ESTIMATE":
             estimate_box = layout.box()
+            available_models = image_depth_model_items(self, context)
+            if not available_models:
+                estimate_box.label(text="No DA3 models installed", icon="ERROR")
+                estimate_box.label(text="Download one in Host > Models")
+            else:
                 estimate_box.prop(self, "snapform_model")
                 estimate_box.prop(self, "snapform_use_ray_pose")
                 process_res = scaled_process_resolution(
@@ -1049,10 +1076,16 @@ class ConvertImageToDepthPlane(bpy.types.Operator):
                     "snapform_process_res_factor",
                     text=f"Resolution Scale ({process_res} px)",
                 )
-        layout.prop(self, "snapform_remove_background")
+        settings_header, settings_body = layout.panel(
-        layout.prop(self, "snapform_displacement_via")
+            "snapform_depth_plane_options",
-        layout.prop(self, "snapform_unshaded")
+            default_closed=True,
-        layout.prop(self, "snapform_subdivision_level")
+        )
+        settings_header.label(text="Options")
+        if settings_body is not None:
+            settings_body.prop(self, "snapform_displacement_via")
+            settings_body.prop(self, "snapform_subdivision_level")
+            settings_body.prop(self, "snapform_remove_background")
+            settings_body.prop(self, "snapform_unshaded")
         layout.prop(self, "snapform_delete_original")
 
     def execute(self, context):

src/snapform/operators.py

@@ -34,7 +34,7 @@ def open_job_directory(result_directory):
 @dataclass
 class ActiveJob:
     kind: str
-    process: subprocess.Popen
+    process: subprocess.Popen | None
     log_file: object
     log_path: Path
     status_path: Path
@@ -50,6 +50,7 @@ class ActiveJob:
     remove_background: bool = False
     temporary_input_path: Path | None = None
     cancelled: bool = False
+    uses_host: bool = False
 
 
 _active_job = None
@@ -65,7 +66,12 @@ def cancel_active_job():
     if _active_job is None:
         return
     _active_job.cancelled = True
-    if _active_job.process.poll() is None:
+    if _active_job.uses_host:
+        try:
+            runtime.cancel_host_job(_active_job.status_path.stem)
+        except RuntimeError:
+            pass
+    elif _active_job.process.poll() is None:
         _active_job.process.terminate()
 
 
@@ -74,6 +80,7 @@ def close_active_job():
 
     if _active_job is None:
         return
+    if _active_job.log_file is not None:
         _active_job.log_file.close()
     if _active_job.temporary_input_path is not None:
         try:
@@ -171,8 +178,18 @@ class RunJob(bpy.types.Operator):
             self.report({"ERROR"}, str(error))
             return {"CANCELLED"}
 
-        log_file = paths["log"].open("w", encoding="utf-8")
+        uses_host = isinstance(command, runtime.HostJobRequest)
+        log_file = None
+        log_path = paths["log"]
         try:
+            if uses_host:
+                runtime.submit_host_job(command)
+                process = None
+                log_path = runtime.host_log_path()
+            else:
+                if self.job_type == "INSTALL_RUNTIME":
+                    runtime.stop_host()
+                log_file = paths["log"].open("w", encoding="utf-8")
                 process = subprocess.Popen(
                     command,
                     stdout=log_file,
@@ -180,7 +197,8 @@ class RunJob(bpy.types.Operator):
                     cwd=runtime.storage_root(),
                     creationflags=runtime.creation_flags(),
                 )
-        except OSError as error:
+        except (OSError, RuntimeError) as error:
+            if log_file is not None:
                 log_file.close()
             self.report({"ERROR"}, f"Unable to start task: {error}")
             return {"CANCELLED"}
@@ -189,7 +207,7 @@ class RunJob(bpy.types.Operator):
             kind=self.job_type,
             process=process,
             log_file=log_file,
-            log_path=paths["log"],
+            log_path=log_path,
             status_path=paths["status"],
             result_directory=paths["result"],
             export_format=(
@@ -215,6 +233,7 @@ class RunJob(bpy.types.Operator):
                 and self.snapform_delete_input
                 else None
             ),
+            uses_host=uses_host,
         )
         settings.status = self._starting_message(settings)
         settings.progress = 0.0
@@ -274,7 +293,7 @@ class RunJob(bpy.types.Operator):
             if not runtime.model_ready(BEN2_MODEL.key):
                 raise RuntimeError(f"Download {BEN2_MODEL.label} first")
             return (
-                runtime.remove_background_command(
+                runtime.remove_background_request(
                     settings,
                     input_path,
                     paths["status"],
@@ -311,7 +330,7 @@ class RunJob(bpy.types.Operator):
                 "PLY export requires a DA3 Giant or Nested Giant model"
             )
         return (
-            runtime.reconstruct_command(
+            runtime.reconstruct_request(
                 settings,
                 export_format,
                 paths["status"],
@@ -372,6 +391,16 @@ class RunJob(bpy.types.Operator):
             return {"CANCELLED"}
 
         self._update_status(context, job)
+        if job.uses_host:
+            status = runtime.read_status(job.status_path) or {}
+            state = status.get("state")
+            if state == "succeeded":
+                return self._finish(context, job, 0)
+            if state in {"failed", "cancelled"}:
+                return self._finish(context, job, 1)
+            if runtime.host_status().get("state") in {"ERROR", "STOPPED"}:
+                return self._finish(context, job, 1)
+            return {"RUNNING_MODAL"}
         return_code = job.process.poll()
         if return_code is None:
             return {"RUNNING_MODAL"}
@@ -389,6 +418,7 @@ class RunJob(bpy.types.Operator):
     def _finish(self, context, job, return_code):
         settings = context.scene.snapform_settings
         self._remove_timer(context)
+        if job.log_file is not None:
             job.log_file.close()
 
         if job.cancelled:
@@ -403,9 +433,14 @@ class RunJob(bpy.types.Operator):
         if return_code != 0 or error:
             settings.status = "Task failed"
             settings.progress = 0.0
+            fallback_error = (
+                "AI Host task failed"
+                if job.uses_host
+                else "Task subprocess failed"
+            )
             self.report(
                 {"ERROR"},
-                f"{error or 'Task subprocess failed'}; log: {job.log_path}",
+                f"{error or fallback_error}; log: {job.log_path}",
             )
             close_active_job()
             self._redraw_ui(context, force=True)
@@ -559,3 +594,46 @@ class CancelJob(bpy.types.Operator):
     def execute(self, _context):
         cancel_active_job()
         return {"FINISHED"}
+
+
+class ControlHost(bpy.types.Operator):
+    bl_idname = "snapform.control_host"
+    bl_label = "Control SnapForm AI Host"
+
+    action: bpy.props.EnumProperty(
+        items=(
+            ("START", "Start", "Start or connect to the shared AI Host"),
+            ("CLEAR", "Clear Models", "Release cached AI models"),
+            ("RESTART", "Restart", "Restart the shared AI Host"),
+            ("STOP", "Stop", "Stop the shared AI Host"),
+            ("LOG", "Open Log", "Open the AI Host log"),
+        ),
+        options={"HIDDEN", "SKIP_SAVE"},
+    )
+
+    def execute(self, _context):
+        try:
+            if self.action == "START":
+                runtime.enable_host()
+                message = "AI Host is ready"
+            elif self.action == "CLEAR":
+                runtime.clear_host_models()
+                message = "AI model cache is cleared"
+            elif self.action == "RESTART":
+                runtime.restart_host()
+                message = "AI Host is restarting"
+            elif self.action == "STOP":
+                runtime.stop_host()
+                message = "AI Host is stopping"
+            else:
+                log_path = runtime.host_log_path()
+                target = log_path if log_path.is_file() else log_path.parent
+                result = bpy.ops.wm.path_open(filepath=str(target))
+                if "FINISHED" not in result:
+                    raise RuntimeError("Unable to open the AI Host log")
+                return {"FINISHED"}
+        except (OSError, RuntimeError) as error:
+            self.report({"ERROR"}, str(error))
+            return {"CANCELLED"}
+        self.report({"INFO"}, message)
+        return {"FINISHED"}

src/snapform/panel.py

@@ -1,21 +1,22 @@
 import bpy
 
-from .models import BEN2_MODEL, MODELS
+from .models import BEN2_MODEL, MODELS, selected_model_key
 from .operators import active_job
+from . import runtime
 from .runtime import model_ready, runtime_ready
 
 
-class RuntimeModelsPanel(bpy.types.Panel):
+class HostPanel(bpy.types.Panel):
-    bl_label = "Setup"
+    bl_label = "Host"
-    bl_idname = "SNAPFORM_PT_runtime_models"
+    bl_idname = "SNAPFORM_PT_host"
     bl_space_type = "VIEW_3D"
     bl_region_type = "UI"
     bl_category = "SnapForm"
+    bl_order = 0
 
     def draw(self, _context):
         layout = self.layout
-        running = active_job() is not None
+        running = active_job() is not None or runtime.host_busy()
-
         if not runtime_ready():
             layout.label(text="Runtime is not installed", icon="ERROR")
             install_row = layout.row()
@@ -26,6 +27,101 @@ class RuntimeModelsPanel(bpy.types.Panel):
                 icon="IMPORT",
             )
             operator.job_type = "INSTALL_RUNTIME"
+            return
+
+        status = runtime.host_status()
+        state = status.get("state", "STOPPED")
+        icons = {
+            "READY": "RADIOBUT_ON",
+            "BUSY": "TIME",
+            "STARTING": "TIME",
+            "ERROR": "ERROR",
+            "UNAVAILABLE": "ERROR",
+            "STOPPED": "PAUSE",
+        }
+        row = layout.row(align=True)
+        if state in {"READY", "BUSY"}:
+            row.label(
+                text=(
+                    f"Running (Clients: {status.get('clients', 0)})"
+                    if state == "READY"
+                    else "Running (Busy)"
+                ),
+                icon=icons[state],
+            )
+            if state == "READY":
+                controls = row.row(align=True)
+                controls.enabled = status.get("clients", 0) <= 1
+                operator = controls.operator(
+                    "snapform.control_host",
+                    text="",
+                    icon="FILE_REFRESH",
+                )
+                operator.action = "RESTART"
+                operator = controls.operator(
+                    "snapform.control_host",
+                    text="",
+                    icon="PAUSE",
+                )
+                operator.action = "STOP"
+        elif state == "STOPPED":
+            operator = row.operator(
+                "snapform.control_host",
+                text="Start",
+                icon="PLAY",
+            )
+            operator.action = "START"
+        else:
+            row.label(
+                text=status.get("message", state.title()),
+                icon=icons.get(state, "QUESTION"),
+            )
+            if state == "ERROR":
+                operator = row.operator(
+                    "snapform.control_host",
+                    text="Start",
+                    icon="PLAY",
+                )
+                operator.action = "START"
+
+        if state in {"READY", "BUSY"}:
+            cached = status.get("cached_models", {})
+            cached_names = [name for name in cached.values() if name]
+            cached_row = layout.row(align=True)
+            cached_row.label(
+                text=(
+                    f"Cached: {', '.join(cached_names)}"
+                    if cached_names
+                    else "Cached: None"
+                )
+            )
+            clear = cached_row.row(align=True)
+            clear.enabled = state == "READY"
+            operator = clear.operator(
+                "snapform.control_host",
+                text="",
+                icon="TRASH",
+            )
+            operator.action = "CLEAR"
+
+
+class ModelsPanel(bpy.types.Panel):
+    bl_label = "Models"
+    bl_idname = "SNAPFORM_PT_models"
+    bl_parent_id = HostPanel.bl_idname
+    bl_space_type = "VIEW_3D"
+    bl_region_type = "UI"
+    bl_category = "SnapForm"
+    bl_options = {"DEFAULT_CLOSED"}
+    bl_order = 0
+
+    @classmethod
+    def poll(cls, _context):
+        return runtime_ready()
+
+    def draw(self, _context):
+        layout = self.layout
+        running = active_job() is not None or runtime.host_busy()
 
         for model in MODELS.values():
             self._draw_model(layout, model, running)
@@ -55,6 +151,24 @@ class DebugPanel(bpy.types.Panel):
     bl_region_type = "UI"
     bl_category = "SnapForm"
     bl_options = {"DEFAULT_CLOSED"}
+    bl_order = 1
+
+    def draw(self, context):
+        layout = self.layout
+        settings = context.scene.snapform_settings
+
+        layout.prop(settings, "device")
+
+
+class Da3DebugPanel(bpy.types.Panel):
+    bl_label = "DA3 Reconstruction"
+    bl_idname = "SNAPFORM_PT_debug_da3"
+    bl_parent_id = DebugPanel.bl_idname
+    bl_space_type = "VIEW_3D"
+    bl_region_type = "UI"
+    bl_category = "SnapForm"
+    bl_options = {"DEFAULT_CLOSED"}
+    bl_order = 0
 
     def draw(self, context):
         layout = self.layout
@@ -82,7 +196,6 @@ class DebugPanel(bpy.types.Panel):
             )
         ):
             layout.prop(settings, "video_frames")
-        layout.prop(settings, "device")
         layout.prop(settings, "process_res")
         layout.prop(settings, "process_res_method")
         layout.prop(settings, "use_ray_pose")
@@ -100,7 +213,11 @@ class DebugPanel(bpy.types.Panel):
             ply_box.prop(settings, "gs_views_interval")
 
         export_row = layout.row()
-        export_row.enabled = active_job() is None
+        export_row.enabled = (
+            active_job() is None
+            and not runtime.host_busy()
+            and model_ready(selected_model_key(settings))
+        )
         operator = export_row.operator(
             "snapform.run_job",
             text="Export",
@@ -108,11 +225,28 @@ class DebugPanel(bpy.types.Panel):
         )
         operator.job_type = "RECONSTRUCT"
 
-        background_box = layout.box()
+
-        background_box.label(text="BEN2 Background Removal")
+class Ben2DebugPanel(bpy.types.Panel):
-        background_box.prop(settings, "background_input_path")
+    bl_label = "BEN2 Background Removal"
-        background_row = background_box.row()
+    bl_idname = "SNAPFORM_PT_debug_ben2"
-        background_row.enabled = active_job() is None
+    bl_parent_id = DebugPanel.bl_idname
+    bl_space_type = "VIEW_3D"
+    bl_region_type = "UI"
+    bl_category = "SnapForm"
+    bl_options = {"DEFAULT_CLOSED"}
+    bl_order = 1
+
+    def draw(self, context):
+        layout = self.layout
+        settings = context.scene.snapform_settings
+
+        layout.prop(settings, "background_input_path")
+        background_row = layout.row()
+        background_row.enabled = (
+            active_job() is None
+            and not runtime.host_busy()
+            and model_ready(BEN2_MODEL.key)
+        )
         operator = background_row.operator(
             "snapform.run_job",
             text="Remove Background",

src/snapform/preferences.py

@@ -53,3 +53,10 @@ class SnapFormPreferences(bpy.types.AddonPreferences):
                 icon="IMPORT",
             )
             operator.job_type = "INSTALL_RUNTIME"
+        layout.separator()
+        operator = layout.operator(
+            "snapform.control_host",
+            text="Open Host Log",
+            icon="TEXT",
+        )
+        operator.action = "LOG"

src/snapform/runtime.py

@@ -1,10 +1,16 @@
 import json
 import os
+import secrets
 import shutil
+import socket
 import subprocess
+import time
 import uuid
+from dataclasses import dataclass
 from datetime import datetime
 from pathlib import Path
+from urllib.error import HTTPError, URLError
+from urllib.request import ProxyHandler, Request, build_opener
 
 import bpy
 
@@ -13,10 +19,40 @@ from .models import BEN2_MODEL, get_downloadable_model, selected_model_key
 
 DA3_COMMIT = "3d835ec1a5802d64a8b8b15f817a1ab54809bfe4"
 BEN2_COMMIT = "2c99a5da477b5523585bfa5c893888a6e818a8f6"
-RUNTIME_SCHEMA = 3
+RUNTIME_SCHEMA = 4
+HOST_SCHEMA = 2
+HOST_ADDRESS = "127.0.0.1"
 DEFAULT_STORAGE_ROOT = Path.home() / ".snapform"
 
 
+@dataclass(frozen=True)
+class HostJobRequest:
+    payload: dict
+
+
+@dataclass
+class HostConnection:
+    port: int
+    token: str
+    pid: int
+    instance_id: str
+
+
+@dataclass
+class LaunchedHost:
+    process: subprocess.Popen
+    log_file: object
+
+
+_host_connection = None
+_launched_host = None
+_registered_instance_id = None
+_client_id = uuid.uuid4().hex
+_host_auto_start = False
+_host_snapshot = {"state": "STOPPED", "message": "Stopped"}
+_local_opener = build_opener(ProxyHandler({}))
+
+
 def preferences():
     addon = bpy.context.preferences.addons.get(__package__)
     return addon.preferences if addon is not None else None
@@ -124,6 +160,8 @@ def install_runtime_command(status_path):
         executable,
         "--runtime-dir",
         str(runtime_directory()),
+        "--models-dir",
+        str(models_directory()),
         "--backend",
         backend.lower(),
         "--da3-commit",
@@ -135,6 +173,419 @@ def install_runtime_command(status_path):
     ]
 
 
+def host_log_path():
+    return storage_root() / "ai-host.log"
+
+
+def host_discovery_path():
+    return storage_root() / "host.json"
+
+
+def host_lock_path():
+    return storage_root() / "host.lock"
+
+
+def _available_host_port():
+    with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as listener:
+        listener.bind((HOST_ADDRESS, 0))
+        return listener.getsockname()[1]
+
+
+def _host_url(connection, path):
+    return f"http://{HOST_ADDRESS}:{connection.port}{path}"
+
+
+def _host_http_request(connection, method, path, payload=None, timeout=2.0):
+    data = None
+    headers = {"X-SnapForm-Token": connection.token}
+    if payload is not None:
+        data = json.dumps(payload).encode("utf-8")
+        headers["Content-Type"] = "application/json"
+    request = Request(
+        _host_url(connection, path),
+        data=data,
+        headers=headers,
+        method=method,
+    )
+    try:
+        with _local_opener.open(request, timeout=timeout) as response:
+            return json.loads(response.read().decode("utf-8"))
+    except HTTPError as error:
+        try:
+            detail = json.loads(error.read().decode("utf-8")).get("detail")
+        except (json.JSONDecodeError, UnicodeDecodeError):
+            detail = None
+        raise RuntimeError(detail or f"AI Host returned HTTP {error.code}")
+    except (OSError, URLError) as error:
+        raise RuntimeError(f"Unable to reach AI Host: {error}")
+
+
+def _read_host_discovery():
+    try:
+        payload = json.loads(host_discovery_path().read_text(encoding="utf-8"))
+        return HostConnection(
+            port=int(payload["port"]),
+            token=str(payload["token"]),
+            pid=int(payload["pid"]),
+            instance_id=str(payload["instance_id"]),
+        )
+    except (KeyError, TypeError, ValueError, OSError, json.JSONDecodeError):
+        return None
+
+
+def _check_host(connection, timeout=0.3):
+    health = _host_http_request(
+        connection,
+        "GET",
+        "/health",
+        timeout=timeout,
+    )
+    if health.get("host_schema") != HOST_SCHEMA:
+        raise RuntimeError("AI Host protocol is out of date")
+    if health.get("instance_id") != connection.instance_id:
+        raise RuntimeError("AI Host discovery is stale")
+    return health
+
+
+def _discover_live_host(timeout=0.3):
+    connection = _read_host_discovery()
+    if connection is None:
+        return None, None
+    try:
+        return connection, _check_host(connection, timeout=timeout)
+    except RuntimeError:
+        return None, None
+
+
+def start_host():
+    global _launched_host
+
+    if _launched_host is not None and _launched_host.process.poll() is None:
+        return
+    _close_launched_host()
+    port = _available_host_port()
+    token = secrets.token_urlsafe(32)
+    instance_id = uuid.uuid4().hex
+    log_file = host_log_path().open("a", encoding="utf-8")
+    command = [
+        str(runtime_python()),
+        "-u",
+        str(Path(__file__).with_name("ai_host.py")),
+        "--host",
+        HOST_ADDRESS,
+        "--port",
+        str(port),
+        "--token",
+        token,
+        "--instance-id",
+        instance_id,
+        "--discovery",
+        str(host_discovery_path()),
+        "--lock",
+        str(host_lock_path()),
+    ]
+    try:
+        process = subprocess.Popen(
+            command,
+            stdout=log_file,
+            stderr=subprocess.STDOUT,
+            cwd=storage_root(),
+            creationflags=creation_flags(),
+        )
+    except OSError:
+        log_file.close()
+        raise
+    _launched_host = LaunchedHost(process, log_file)
+
+
+def ensure_host(timeout=60.0):
+    global _host_auto_start, _host_connection, _host_snapshot
+
+    _host_auto_start = True
+    deadline = time.monotonic() + timeout
+    last_launch = 0.0
+    starting = False
+    while time.monotonic() < deadline:
+        connection, _health = _discover_live_host(timeout=0.5)
+        if connection is not None:
+            _host_connection = connection
+            _register_host_client(connection)
+            health = _check_host(connection, timeout=0.5)
+            _host_snapshot = _format_host_snapshot(connection, health)
+            _redraw_host_ui()
+            return connection
+        if not starting:
+            _host_snapshot = {"state": "STARTING", "message": "Starting"}
+            _redraw_host_ui()
+            starting = True
+        _close_finished_launched_host()
+        if (
+            _launched_host is None
+            and time.monotonic() - last_launch >= 1.0
+        ):
+            start_host()
+            last_launch = time.monotonic()
+        time.sleep(0.05)
+    raise RuntimeError(f"AI Host did not become ready; log: {host_log_path()}")
+
+
+def submit_host_job(request):
+    connection = ensure_host()
+    payload = dict(request.payload)
+    payload["client_id"] = _client_id
+    _host_http_request(
+        connection,
+        "POST",
+        "/jobs",
+        payload=payload,
+        timeout=5.0,
+    )
+    return connection
+
+
+def stop_host():
+    global _host_auto_start, _host_connection, _host_snapshot
+    global _registered_instance_id
+
+    connection = _host_connection
+    if connection is None:
+        connection, _health = _discover_live_host(timeout=0.5)
+    if connection is None:
+        _close_launched_host(terminate=True)
+        _host_auto_start = False
+        _host_snapshot = {"state": "STOPPED", "message": "Stopped"}
+        return
+    _register_host_client(connection)
+    _host_http_request(
+        connection,
+        "POST",
+        "/shutdown",
+        payload={"client_id": _client_id},
+        timeout=2.0,
+    )
+    _host_auto_start = False
+    deadline = time.monotonic() + 10.0
+    while time.monotonic() < deadline:
+        try:
+            _check_host(connection, timeout=0.2)
+        except RuntimeError:
+            break
+        time.sleep(0.05)
+    else:
+        raise RuntimeError("AI Host did not stop")
+    _host_connection = None
+    _registered_instance_id = None
+    _close_finished_launched_host()
+    _host_snapshot = {"state": "STOPPED", "message": "Stopped"}
+
+
+def restart_host():
+    global _host_auto_start, _host_connection, _registered_instance_id
+    global _host_snapshot
+
+    connection = ensure_host()
+    _host_http_request(
+        connection,
+        "POST",
+        "/shutdown",
+        payload={"client_id": _client_id},
+        timeout=2.0,
+    )
+    _host_auto_start = True
+    _host_connection = None
+    _registered_instance_id = None
+    _host_snapshot = {"state": "STARTING", "message": "Restarting"}
+
+
+def enable_host():
+    global _host_auto_start, _host_connection, _host_snapshot
+
+    _host_auto_start = True
+    connection, health = _discover_live_host(timeout=0.2)
+    if connection is not None:
+        _host_connection = connection
+        _register_host_client(connection)
+        _host_snapshot = _format_host_snapshot(connection, health)
+        _redraw_host_ui()
+        return
+    start_host()
+    _host_snapshot = {"state": "STARTING", "message": "Starting"}
+    _redraw_host_ui()
+
+
+def clear_host_models():
+    connection = ensure_host()
+    return _host_http_request(
+        connection,
+        "POST",
+        "/models/clear",
+        timeout=5.0,
+    )
+
+
+def cancel_host_job(job_id):
+    connection = ensure_host()
+    return _host_http_request(
+        connection,
+        "DELETE",
+        f"/jobs/{job_id}",
+        payload={"client_id": _client_id},
+        timeout=2.0,
+    )
+
+
+def detach_host_client():
+    global _host_connection, _registered_instance_id
+
+    connection = _host_connection
+    if connection is not None and _registered_instance_id is not None:
+        try:
+            _host_http_request(
+                connection,
+                "DELETE",
+                f"/clients/{_client_id}",
+                timeout=0.5,
+            )
+        except RuntimeError:
+            pass
+    _host_connection = None
+    _registered_instance_id = None
+    _close_launched_host(terminate=False)
+
+
+def schedule_host_start():
+    timers = getattr(bpy.app, "timers", None)
+    if timers is None:
+        return
+    if timers.is_registered(_service_host_timer):
+        return
+    timers.register(_service_host_timer, first_interval=1.0, persistent=True)
+
+
+def unschedule_host():
+    timers = getattr(bpy.app, "timers", None)
+    if timers is None:
+        return
+    if timers.is_registered(_service_host_timer):
+        timers.unregister(_service_host_timer)
+
+
+def host_status():
+    return dict(_host_snapshot)
+
+
+def host_busy():
+    return _host_snapshot.get("state") == "BUSY"
+
+
+def _service_host_timer():
+    global _host_connection, _host_snapshot, _registered_instance_id
+
+    if not runtime_ready():
+        _host_snapshot = {
+            "state": "UNAVAILABLE",
+            "message": "Runtime is not installed",
+        }
+        return 2.0
+    connection, health = _discover_live_host(timeout=0.2)
+    if connection is None:
+        _host_connection = None
+        _registered_instance_id = None
+        _close_finished_launched_host()
+        if not _host_auto_start:
+            if _host_snapshot.get("state") != "STOPPED":
+                _host_snapshot = {"state": "STOPPED", "message": "Stopped"}
+                _redraw_host_ui()
+            return 2.0
+        try:
+            start_host()
+        except OSError as error:
+            _host_snapshot = {"state": "ERROR", "message": str(error)}
+            return 2.0
+        _host_snapshot = {"state": "STARTING", "message": "Starting"}
+        _redraw_host_ui()
+        return 1.0
+    try:
+        _host_connection = connection
+        if _registered_instance_id == connection.instance_id:
+            try:
+                _host_http_request(
+                    connection,
+                    "POST",
+                    f"/clients/{_client_id}/heartbeat",
+                    timeout=0.5,
+                )
+            except RuntimeError:
+                _registered_instance_id = None
+        _register_host_client(connection)
+        health = _check_host(connection, timeout=0.5)
+        _host_snapshot = _format_host_snapshot(connection, health)
+    except RuntimeError as error:
+        _host_connection = None
+        _registered_instance_id = None
+        _host_snapshot = {"state": "ERROR", "message": str(error)}
+    _redraw_host_ui()
+    return 5.0
+
+
+def _register_host_client(connection):
+    global _registered_instance_id
+
+    if _registered_instance_id == connection.instance_id:
+        return
+    label = f"Blender {getattr(bpy.app, 'version_string', '')}".strip()
+    _host_http_request(
+        connection,
+        "POST",
+        "/clients",
+        payload={"client_id": _client_id, "label": label},
+        timeout=2.0,
+    )
+    _registered_instance_id = connection.instance_id
+
+
+def _format_host_snapshot(connection, health):
+    state = "BUSY" if health.get("busy") else "READY"
+    return {
+        **health,
+        "state": state,
+        "message": "Busy" if state == "BUSY" else "Ready",
+        "port": connection.port,
+    }
+
+
+def _redraw_host_ui():
+    try:
+        from .status_bar import redraw_ui
+
+        redraw_ui(bpy.context, force=True)
+    except (AttributeError, RuntimeError):
+        pass
+
+
+def _close_finished_launched_host():
+    if _launched_host is None or _launched_host.process.poll() is None:
+        return
+    _close_launched_host()
+
+
+def _close_launched_host(terminate=False):
+    global _launched_host
+
+    launched = _launched_host
+    _launched_host = None
+    if launched is None:
+        return
+    if terminate and launched.process.poll() is None:
+        launched.process.terminate()
+        try:
+            launched.process.wait(timeout=3.0)
+        except subprocess.TimeoutExpired:
+            launched.process.kill()
+            launched.process.wait(timeout=3.0)
+    launched.log_file.close()
+
+
 def download_model_command(model_key, status_path):
     model = get_downloadable_model(model_key)
     command = [
@@ -153,31 +604,29 @@ def download_model_command(model_key, status_path):
     return command
 
 
-def remove_background_command(
+def remove_background_request(
     settings,
     input_path,
     status_path,
     result_directory,
     model_key,
 ):
-    return [
+    return HostJobRequest(
-        str(runtime_python()),
+        {
-        str(Path(__file__).with_name("worker.py")),
+            "job_id": Path(status_path).stem,
-        "remove-background",
+            "command": "remove-background",
-        "--input",
+            "parameters": {
-        str(input_path),
+                "input": str(input_path),
-        "--model-dir",
+                "model_dir": str(model_directory(model_key)),
-        str(model_directory(model_key)),
+                "output_dir": str(result_directory),
-        "--output-dir",
+                "status": str(status_path),
-        str(result_directory),
+                "device": settings.device.lower(),
-        "--status",
+            },
-        str(status_path),
+        }
-        "--device",
+    )
-        settings.device.lower(),
-    ]
 
 
-def reconstruct_command(
+def reconstruct_request(
     settings,
     export_format,
     status_path,
@@ -201,53 +650,39 @@ def reconstruct_command(
     snapform_selected_ray_pose = (
         settings.use_ray_pose if use_ray_pose is None else use_ray_pose
     )
-    command = [
+    parameters = {
-        str(runtime_python()),
+        "input": str(
-        str(Path(__file__).with_name("worker.py")),
-        "reconstruct",
-        "--input",
-        str(
             Path(input_path)
             if input_path is not None
             else Path(bpy.path.abspath(settings.input_path))
         ),
-        "--model-dir",
+        "model_dir": str(model_directory(snapform_selected_key)),
-        str(model_directory(snapform_selected_key)),
+        "output_dir": str(result_directory),
-        "--output-dir",
+        "export_format": export_format,
-        str(result_directory),
+        "status": str(status_path),
-        "--export-format",
+        "process_res": snapform_selected_process_res,
-        export_format,
+        "process_res_method": snapform_selected_process_res_method,
-        "--status",
+        "conf_thresh_percentile": settings.conf_thresh_percentile,
-        str(status_path),
+        "num_max_points": settings.num_max_points,
-        "--process-res",
+        "show_cameras": 1 if settings.show_cameras else 0,
-        str(snapform_selected_process_res),
+        "gs_views_interval": settings.gs_views_interval,
-        "--process-res-method",
+        "video_frames": settings.video_frames,
-        snapform_selected_process_res_method,
+        "device": settings.device.lower(),
-        "--conf-thresh-percentile",
+        "use_ray_pose": 1 if snapform_selected_ray_pose else 0,
-        str(settings.conf_thresh_percentile),
+        "ref_view_strategy": settings.ref_view_strategy,
-        "--num-max-points",
+        "remove_background_model_dir": "",
-        str(settings.num_max_points),
+    }
-        "--show-cameras",
-        "1" if settings.show_cameras else "0",
-        "--gs-views-interval",
-        str(settings.gs_views_interval),
-        "--video-frames",
-        str(settings.video_frames),
-        "--device",
-        settings.device.lower(),
-        "--use-ray-pose",
-        "1" if snapform_selected_ray_pose else "0",
-        "--ref-view-strategy",
-        settings.ref_view_strategy,
-    ]
     if remove_background:
-        command.extend(
+        parameters["remove_background_model_dir"] = str(
-            [
+            model_directory(BEN2_MODEL.key)
-                "--remove-background-model-dir",
+        )
-                str(model_directory(BEN2_MODEL.key)),
+    return HostJobRequest(
-            ]
+        {
+            "job_id": Path(status_path).stem,
+            "command": "reconstruct",
+            "parameters": parameters,
+        }
     )
-    return command
 
 
 def creation_flags():

src/snapform/worker.py

@@ -23,10 +23,27 @@ SCENE_FILENAME = "scene.glb"
 EXPORT_FORMATS = ("glb", "gs_ply", "depth_vis", "mini_npz")
 
 
-def write_status(path, progress, message, error=None):
+class TaskCancelled(RuntimeError):
+    pass
+
+
+def check_cancelled(cancel_check):
+    if cancel_check is not None and cancel_check():
+        raise TaskCancelled("Task cancelled")
+
+
+def write_status(path, progress, message, error=None, state=None):
     payload = {"progress": progress, "message": message}
     if error:
         payload["error"] = error
+    if state is None:
+        if error:
+            state = "failed"
+        elif progress >= 1.0:
+            state = "succeeded"
+        else:
+            state = "running"
+    payload["state"] = state
     Path(path).write_text(
         json.dumps(payload, ensure_ascii=False),
         encoding="utf-8",
@@ -42,7 +59,13 @@ def collect_images(directory):
     )
 
 
-def extract_video_frames(video_path, output_directory, maximum_frames, cv2):
+def extract_video_frames(
+    video_path,
+    output_directory,
+    maximum_frames,
+    cv2,
+    cancel_check=None,
+):
     capture = cv2.VideoCapture(str(video_path))
     if not capture.isOpened():
         raise RuntimeError(f"Unable to open video: {video_path}")
@@ -66,6 +89,7 @@ def extract_video_frames(video_path, output_directory, maximum_frames, cv2):
     image_paths = []
     try:
         for output_index, frame_index in enumerate(frame_indices):
+            check_cancelled(cancel_check)
             capture.set(cv2.CAP_PROP_POS_FRAMES, int(frame_index))
             success, frame = capture.read()
             if not success:
@@ -82,7 +106,8 @@ def extract_video_frames(video_path, output_directory, maximum_frames, cv2):
     return image_paths
 
 
-def prepare_input(input_path, video_frames, temporary_root):
+def prepare_input(input_path, video_frames, temporary_root, cancel_check=None):
+    check_cancelled(cancel_check)
     input_path = Path(input_path)
     suffix = input_path.suffix.lower()
     if suffix in IMAGE_SUFFIXES:
@@ -102,6 +127,7 @@ def prepare_input(input_path, video_frames, temporary_root):
             temporary_directory,
             video_frames,
             cv2,
+            cancel_check=cancel_check,
         )
     except Exception:
         shutil.rmtree(temporary_directory, ignore_errors=True)
@@ -203,19 +229,33 @@ def select_device(requested, torch):
     return torch.device("cuda" if torch.cuda.is_available() else "cpu")
 
 
-def infer_foreground(input_path, model_dir, device, torch):
+def infer_foreground(
-    from ben2 import BEN_Base
+    input_path,
+    model_dir,
+    device,
+    torch,
+    model_cache=None,
+    cancel_check=None,
+):
     from PIL import Image
 
     input_path = Path(input_path)
+    check_cancelled(cancel_check)
     if input_path.suffix.lower() not in IMAGE_SUFFIXES:
         raise RuntimeError("BEN2 requires a JPG, JPEG, PNG, or WEBP image")
+    owns_model = model_cache is None
+    if owns_model:
+        from ben2 import BEN_Base
+
         load_started = time.perf_counter()
         model = BEN_Base.from_pretrained(model_dir)
         model = model.to(device=device)
         model.eval()
         load_ms = (time.perf_counter() - load_started) * 1000.0
+    else:
+        model, load_ms = model_cache.get_ben2(model_dir, device)
 
+    check_cancelled(cancel_check)
     with Image.open(input_path) as opened_image:
         image = opened_image.convert("RGB")
     if device.type == "cuda":
@@ -229,17 +269,23 @@ def infer_foreground(input_path, model_dir, device, torch):
     if device.type == "cuda":
         torch.cuda.synchronize()
     inference_ms = (time.perf_counter() - inference_started) * 1000.0
+    if owns_model:
         del model
         if device.type == "cuda":
             torch.cuda.empty_cache()
+    check_cancelled(cancel_check)
     return foreground, load_ms, inference_ms
 
 
-def remove_background(args):
+def remove_background(args, model_cache=None, cancel_check=None):
     total_started = time.perf_counter()
     write_status(args.status, 0.02, "Starting BEN2 Runtime")
+    check_cancelled(cancel_check)
 
+    if model_cache is None:
         import torch
+    else:
+        torch = model_cache.torch
 
     output_directory = Path(args.output_dir)
     output_directory.mkdir(parents=True, exist_ok=False)
@@ -252,8 +298,11 @@ def remove_background(args):
         args.model_dir,
         device,
         torch,
+        model_cache=model_cache,
+        cancel_check=cancel_check,
     )
 
+    check_cancelled(cancel_check)
     write_status(args.status, 0.92, "Writing transparent PNG")
     output_path = output_directory / "foreground.png"
     foreground.convert("RGBA").save(
@@ -264,6 +313,7 @@ def remove_background(args):
     if not output_path.is_file():
         raise RuntimeError("BEN2 did not write a foreground image")
 
+    check_cancelled(cancel_check)
     total_ms = (time.perf_counter() - total_started) * 1000.0
     message = (
         f"Background removed in {inference_ms:.0f} ms "
@@ -288,10 +338,16 @@ def exported_paths(output_directory, export_format):
     return []
 
 
-def wait_for_exported_paths(output_directory, export_format, timeout=30.0):
+def wait_for_exported_paths(
+    output_directory,
+    export_format,
+    timeout=30.0,
+    cancel_check=None,
+):
     # 上游 NPZ exporter 使用后台线程，因此需要等待文件实际落盘。
     deadline = time.monotonic() + timeout
     while time.monotonic() < deadline:
+        check_cancelled(cancel_check)
         result_paths = exported_paths(output_directory, export_format)
         if result_paths:
             return result_paths
@@ -299,11 +355,15 @@ def wait_for_exported_paths(output_directory, export_format, timeout=30.0):
     return []
 
 
-def reconstruct(args):
+def reconstruct(args, model_cache=None, cancel_check=None):
     write_status(args.status, 0.02, "Starting AI runtime")
+    check_cancelled(cancel_check)
 
+    if model_cache is None:
         import torch
         from depth_anything_3.api import DepthAnything3
+    else:
+        torch = model_cache.torch
     from depth_anything_3.utils.export import export
 
     output_directory = Path(args.output_dir)
@@ -313,9 +373,11 @@ def reconstruct(args):
         args.input,
         args.video_frames,
         output_directory.parent,
+        cancel_check=cancel_check,
     )
 
     try:
+        check_cancelled(cancel_check)
         device = select_device(args.device, torch)
         remove_background_model_dir = getattr(
             args,
@@ -329,7 +391,10 @@ def reconstruct(args):
                 remove_background_model_dir,
                 device,
                 torch,
+                model_cache=model_cache,
+                cancel_check=cancel_check,
             )
+            check_cancelled(cancel_check)
             foreground.convert("RGBA").save(
                 output_directory / "foreground.png",
                 format="PNG",
@@ -342,10 +407,14 @@ def reconstruct(args):
                 f"model load {load_ms:.0f} ms"
             )
         write_status(args.status, 0.15, f"Loading model on {device.type}")
+        if model_cache is None:
             model = DepthAnything3.from_pretrained(args.model_dir)
             model = model.to(device=device)
             model.eval()
+        else:
+            model = model_cache.get_da3(args.model_dir, device)
 
+        check_cancelled(cancel_check)
         write_status(
             args.status,
             0.25,
@@ -372,6 +441,7 @@ def reconstruct(args):
                 use_ray_pose=bool(args.use_ray_pose),
                 ref_view_strategy=args.ref_view_strategy,
             )
+        check_cancelled(cancel_check)
         write_status(args.status, 0.88, f"Exporting {args.export_format}")
         export(
             prediction,
@@ -379,15 +449,18 @@ def reconstruct(args):
             str(output_directory),
             **export_kwargs,
         )
+        check_cancelled(cancel_check)
 
         result_paths = wait_for_exported_paths(
             output_directory,
             args.export_format,
+            cancel_check=cancel_check,
         )
         if not result_paths:
             raise RuntimeError(
                 f"No {args.export_format} files were exported"
             )
+        check_cancelled(cancel_check)
         write_status(
             args.status,
             1.0,

tests/test_addon.py

@@ -343,7 +343,7 @@ class AddonTest(unittest.TestCase):
         self.assertTrue(operator.__annotations__["snapform_use_ray_pose"]["default"])
         self.assertEqual(
             operator.__annotations__["snapform_process_res_factor"]["default"],
-            0.5,
+            1.0,
         )
         self.assertTrue(
             operator.__annotations__["snapform_delete_original"]["default"]
@@ -403,7 +403,7 @@ class AddonTest(unittest.TestCase):
                 '"snapform_process_res_factor",'
             ),
             draw_source.index(
-                'layout.prop(self, "snapform_subdivision_level")'
+                'settings_body.prop(self, "snapform_subdivision_level")'
             ),
         )
         self.assertIn(
@@ -420,17 +420,30 @@ class AddonTest(unittest.TestCase):
             draw_source,
         )
         self.assertIn(
-            'layout.prop(self, "snapform_remove_background")',
+            'default_closed=True',
             draw_source,
         )
+        self.assertIn('settings_header.label(text="Options")', draw_source)
+        self.assertIn('settings_header.label(text="Options")', draw_source)
         self.assertNotIn("layout.separator", draw_source)
-        self.assertNotIn("layout.label", draw_source)
+        setting_fields = (
-        self.assertIn(
+            "snapform_displacement_via",
-            'layout.prop(self, "snapform_unshaded")',
+            "snapform_subdivision_level",
-            draw_source,
+            "snapform_remove_background",
+            "snapform_unshaded",
         )
-        model_items = operator.__annotations__["snapform_model"]["items"]
+        field_positions = [
-        model_items = model_items(None, None)
+            draw_source.index(f'settings_body.prop(self, "{field}")')
+            for field in setting_fields
+        ]
+        self.assertEqual(field_positions, sorted(field_positions))
+        model_items_callback = operator.__annotations__["snapform_model"]["items"]
+        with patch.object(
+            self.addon.runtime,
+            "model_ready",
+            return_value=True,
+        ):
+            model_items = model_items_callback(None, None)
         self.assertEqual(
             [item[0] for item in model_items],
             [
@@ -459,11 +472,33 @@ class AddonTest(unittest.TestCase):
         )
 
         self.addon.image_depth._last_depth_plane_model = "GIANT_1_1"
+        with patch.object(
+            self.addon.runtime,
+            "model_ready",
+            side_effect=lambda key: key == "GIANT_1_1",
+        ):
             result = operator.invoke(context, None)
 
         self.assertIs(result, operator)
         self.assertEqual(operator.snapform_model, "GIANT_1_1")
 
+    def test_conversion_dialog_only_lists_installed_models(self):
+        items_callback = self.addon.ConvertImageToDepthPlane.__annotations__[
+            "snapform_model"
+        ]["items"]
+        with patch.object(
+            self.addon.runtime,
+            "model_ready",
+            side_effect=lambda key: key in {"SMALL", "BASE"},
+        ):
+            items = items_callback(None, None)
+
+        self.assertEqual([item[0] for item in items], ["SMALL", "BASE"])
+        draw_source = inspect.getsource(self.addon.ConvertImageToDepthPlane.draw)
+        self.assertIn('text="No DA3 models installed"', draw_source)
+        self.assertIn('text="Download one in Host > Models"', draw_source)
+        self.assertNotIn("_dialog_requirements_ready", inspect.getsource(self.addon.ConvertImageToDepthPlane))
+
     def test_depth_conversion_progress_is_provider_neutral(self):
         starting_source = inspect.getsource(
             self.addon.RunJob._starting_message
@@ -532,7 +567,11 @@ class AddonTest(unittest.TestCase):
                 self.addon.image_depth,
                 "save_packed_image",
                 return_value=packed_path,
-            ) as save_packed_image:
+            ) as save_packed_image, patch.object(
+                self.addon.runtime,
+                "model_ready",
+                return_value=True,
+            ):
                 result = operator.execute(context)
 
         self.assertEqual(result, {"FINISHED"})
@@ -640,8 +679,10 @@ class AddonTest(unittest.TestCase):
                     return_value=True,
                 ) as model_ready, patch.object(
                     self.addon.runtime,
-                    "reconstruct_command",
+                    "reconstruct_request",
-                    return_value=["worker"],
+                    return_value=self.addon.runtime.HostJobRequest(
+                        {"job_id": "test"}
+                    ),
                 ) as reconstruct:
                     operator.snapform_model_key = snapform_model
                     operator.snapform_use_ray_pose = snapform_ray_pose
@@ -650,7 +691,10 @@ class AddonTest(unittest.TestCase):
                         settings,
                     )
 
-                self.assertEqual(command, ["worker"])
+                self.assertIsInstance(
+                    command,
+                    self.addon.runtime.HostJobRequest,
+                )
                 self.assertEqual(prepared_paths, paths)
                 model_ready.assert_called_once_with(snapform_model)
                 self.assertEqual(reconstruct.call_args.args[1], "mini_npz")
@@ -713,7 +757,7 @@ class AddonTest(unittest.TestCase):
                 with self.assertRaisesRegex(RuntimeError, "BEN2 Base"):
                     operator._prepare_command(None, settings)
 
-    def test_image_depth_background_option_reaches_worker_command(self):
+    def test_image_depth_background_option_reaches_host_request(self):
         runtime = importlib.import_module("snapform.runtime")
         settings = SimpleNamespace(
             input_path="input.png",
@@ -739,7 +783,7 @@ class AddonTest(unittest.TestCase):
             "model_directory",
             side_effect=lambda key: Path(key.lower()),
         ):
-            command = runtime.reconstruct_command(
+            request = runtime.reconstruct_request(
                 settings,
                 "mini_npz",
                 Path("status.json"),
@@ -747,8 +791,10 @@ class AddonTest(unittest.TestCase):
                 remove_background=True,
             )
 
-        option_index = command.index("--remove-background-model-dir")
+        self.assertEqual(
-        self.assertEqual(command[option_index + 1], str(Path("ben2_base")))
+            request.payload["parameters"]["remove_background_model_dir"],
+            str(Path("ben2_base")),
+        )
 
     def test_preferences_use_package_identifier(self):
         self.assertEqual(
@@ -767,6 +813,25 @@ class AddonTest(unittest.TestCase):
             str(runtime.DEFAULT_STORAGE_ROOT),
         )
 
+    def test_ai_host_uses_authenticated_loopback_http(self):
+        runtime = importlib.import_module("snapform.runtime")
+        request_source = inspect.getsource(runtime._host_http_request)
+        start_source = inspect.getsource(runtime.start_host)
+
+        self.assertEqual(runtime.HOST_ADDRESS, "127.0.0.1")
+        self.assertIn('"X-SnapForm-Token"', request_source)
+        self.assertIn('"/health"', inspect.getsource(runtime._check_host))
+        self.assertIn('with_name("ai_host.py")', start_source)
+        self.assertIn("ProxyHandler({})", inspect.getsource(runtime))
+        self.assertIn('return storage_root() / "host.json"', inspect.getsource(runtime.host_discovery_path))
+
+    def test_host_jobs_complete_from_status_state(self):
+        source = inspect.getsource(self.addon.RunJob.modal)
+
+        self.assertIn("if job.uses_host:", source)
+        self.assertIn('state == "succeeded"', source)
+        self.assertIn('{"failed", "cancelled"}', source)
+
     def test_process_res_method_defaults_to_upper_bound_resize(self):
         process_res = self.addon.SnapFormSettings.__annotations__[
             "process_res"
@@ -780,7 +845,7 @@ class AddonTest(unittest.TestCase):
         self.assertEqual(process_res["default"], 504)
         self.assertEqual(process_res_method["default"], "upper_bound_resize")
 
-    def test_reconstruct_command_preserves_da3_parameter_names(self):
+    def test_reconstruct_request_preserves_da3_parameter_names(self):
         runtime = importlib.import_module("snapform.runtime")
         settings = SimpleNamespace(
             input_path="input.png",
@@ -806,29 +871,30 @@ class AddonTest(unittest.TestCase):
             "model_directory",
             return_value=Path("model"),
         ):
-            command = runtime.reconstruct_command(
+            request = runtime.reconstruct_request(
                 settings,
                 "gs_ply",
                 Path("status.json"),
                 Path("result"),
             )
 
-        self.assertIn("--export-format", command)
+        parameters = request.payload["parameters"]
-        self.assertIn("gs_ply", command)
+        self.assertEqual(request.payload["command"], "reconstruct")
+        self.assertEqual(parameters["export_format"], "gs_ply")
         for parameter in (
-            "--process-res",
+            "process_res",
-            "--process-res-method",
+            "process_res_method",
-            "--conf-thresh-percentile",
+            "conf_thresh_percentile",
-            "--num-max-points",
+            "num_max_points",
-            "--show-cameras",
+            "show_cameras",
-            "--gs-views-interval",
+            "gs_views_interval",
-            "--use-ray-pose",
+            "use_ray_pose",
-            "--ref-view-strategy",
+            "ref_view_strategy",
         ):
             with self.subTest(parameter=parameter):
-                self.assertIn(parameter, command)
+                self.assertIn(parameter, parameters)
 
-    def test_reconstruct_command_accepts_image_depth_overrides(self):
+    def test_reconstruct_request_accepts_image_depth_overrides(self):
         runtime = importlib.import_module("snapform.runtime")
         settings = SimpleNamespace(
             input_path="input.png",
@@ -854,7 +920,7 @@ class AddonTest(unittest.TestCase):
             "model_directory",
             return_value=Path("large-model"),
         ) as snapform_model_directory:
-            snapform_command = runtime.reconstruct_command(
+            request = runtime.reconstruct_request(
                 settings,
                 "mini_npz",
                 Path("status.json"),
@@ -866,15 +932,13 @@ class AddonTest(unittest.TestCase):
             )
 
         snapform_model_directory.assert_called_once_with("BASE")
-        snapform_process_res_index = snapform_command.index("--process-res")
+        parameters = request.payload["parameters"]
-        self.assertEqual(snapform_command[snapform_process_res_index + 1], "252")
+        self.assertEqual(parameters["process_res"], 252)
-        snapform_process_res_method_index = snapform_command.index("--process-res-method")
         self.assertEqual(
-            snapform_command[snapform_process_res_method_index + 1],
+            parameters["process_res_method"],
             "lower_bound_resize",
         )
-        snapform_ray_pose_index = snapform_command.index("--use-ray-pose")
+        self.assertEqual(parameters["use_ray_pose"], 0)
-        self.assertEqual(snapform_command[snapform_ray_pose_index + 1], "0")
 
     def test_export_format_property_includes_all_da3_formats(self):
         export_format = self.addon.SnapFormSettings.__annotations__[
@@ -904,37 +968,139 @@ class AddonTest(unittest.TestCase):
             "DEFAULT_CLOSED",
             self.addon.DebugPanel.bl_options,
         )
+        for panel in (self.addon.Da3DebugPanel, self.addon.Ben2DebugPanel):
+            self.assertEqual(panel.bl_parent_id, self.addon.DebugPanel.bl_idname)
+            self.assertIn("DEFAULT_CLOSED", panel.bl_options)
 
     def test_debug_panel_uses_one_export_button_and_dynamic_settings(self):
-        draw_source = inspect.getsource(
+        debug_source = inspect.getsource(self.addon.DebugPanel.draw)
-            self.addon.DebugPanel.draw
+        da3_source = inspect.getsource(self.addon.Da3DebugPanel.draw)
-        )
+        ben2_source = inspect.getsource(self.addon.Ben2DebugPanel.draw)
 
-        self.assertEqual(draw_source.count("export_row.operator("), 1)
+        self.assertIn('layout.prop(settings, "device")', debug_source)
-        self.assertIn('settings.export_format == "glb"', draw_source)
+        self.assertEqual(da3_source.count("export_row.operator("), 1)
-        self.assertIn('settings.export_format == "gs_ply"', draw_source)
+        self.assertIn('settings.export_format == "glb"', da3_source)
-        self.assertIn('"gs_model" if settings.export_format', draw_source)
+        self.assertIn('settings.export_format == "gs_ply"', da3_source)
-        self.assertIn("layout.prop(settings, model_property)", draw_source)
+        self.assertIn('"gs_model" if settings.export_format', da3_source)
-        self.assertNotIn("model_row", draw_source)
+        self.assertIn("layout.prop(settings, model_property)", da3_source)
-        self.assertIn("export_row.enabled = active_job() is None", draw_source)
+        self.assertNotIn("model_row", da3_source)
-        self.assertNotIn('text="Ready"', draw_source)
+        self.assertIn("and model_ready(selected_model_key(settings))", da3_source)
-        self.assertNotIn('text="Missing"', draw_source)
+        self.assertNotIn('text="Ready"', da3_source)
-        self.assertNotIn("License:", draw_source)
+        self.assertNotIn('text="Missing"', da3_source)
-        self.assertIn('operator.job_type = "RECONSTRUCT"', draw_source)
+        self.assertNotIn("License:", da3_source)
-        self.assertIn('text="BEN2 Background Removal"', draw_source)
+        self.assertIn('operator.job_type = "RECONSTRUCT"', da3_source)
         self.assertIn(
-            'background_box.prop(settings, "background_input_path")',
+            'layout.prop(settings, "background_input_path")',
-            draw_source,
+            ben2_source,
         )
-        self.assertIn('text="Remove Background"', draw_source)
+        self.assertIn('text="Remove Background"', ben2_source)
-        self.assertIn('operator.job_type = "REMOVE_BACKGROUND"', draw_source)
+        self.assertIn("and model_ready(BEN2_MODEL.key)", ben2_source)
+        self.assertIn('operator.job_type = "REMOVE_BACKGROUND"', ben2_source)
 
-    def test_runtime_models_panel_assigns_each_download_target(self):
+    def test_host_panel_exposes_shared_host_state_and_controls(self):
+        self.assertIn(self.addon.HostPanel, self.addon.CLASSES)
+        self.assertIn(self.addon.ControlHost, self.addon.CLASSES)
+        source = inspect.getsource(self.addon.HostPanel.draw)
+
+        self.assertEqual(self.addon.HostPanel.bl_label, "Host")
+        self.assertIn('text="Runtime is not installed"', source)
+        self.assertIn('text="Install Runtime"', source)
+        self.assertIn("runtime.host_status()", source)
+        self.assertIn('f"Running (Clients:', source)
+        self.assertIn('else "Running (Busy)"', source)
+        self.assertIn('"READY": "RADIOBUT_ON"', source)
+        self.assertIn('"BUSY": "TIME"', source)
+        self.assertNotIn('"READY": "PLAY"', source)
+        self.assertNotIn('text=f"Task:', source)
+        self.assertIn('elif state == "STOPPED":', source)
+        self.assertNotIn("Blender Clients:", source)
+        self.assertNotIn("status['port']", source)
+        self.assertIn('icon="TRASH"', source)
+        self.assertIn('icon="FILE_REFRESH"', source)
+        self.assertIn('icon="PAUSE"', source)
+        self.assertNotIn('text="Open Log"', source)
+        preferences_source = inspect.getsource(self.addon.SnapFormPreferences.draw)
+        self.assertIn('text="Open Host Log"', preferences_source)
+
+    def test_host_start_control_does_not_wait_for_model_runtime(self):
+        source = inspect.getsource(self.addon.runtime.enable_host)
+
+        self.assertIn("start_host()", source)
+        self.assertNotIn("ensure_host()", source)
+
+    def test_ai_operation_restarts_a_manually_stopped_host(self):
+        runtime = self.addon.runtime
+        connection = runtime.HostConnection(51234, "token", 42, "host")
+        health = {
+            "host_schema": runtime.HOST_SCHEMA,
+            "instance_id": "host",
+            "busy": False,
+            "clients": 1,
+            "cached_models": {},
+        }
+        previous_auto_start = runtime._host_auto_start
+        previous_connection = runtime._host_connection
+        previous_snapshot = runtime._host_snapshot
+        runtime._host_auto_start = False
+        runtime._host_connection = None
+        runtime._host_snapshot = {"state": "STOPPED", "message": "Stopped"}
+        try:
+            with patch.object(
+                runtime,
+                "_discover_live_host",
+                return_value=(connection, health),
+            ), patch.object(
+                runtime,
+                "_register_host_client",
+            ), patch.object(
+                runtime,
+                "_check_host",
+                return_value=health,
+            ), patch.object(runtime, "_redraw_host_ui"):
+                result = runtime.ensure_host(timeout=0.1)
+
+            self.assertIs(result, connection)
+            self.assertTrue(runtime._host_auto_start)
+            self.assertEqual(runtime.host_status()["state"], "READY")
+        finally:
+            runtime._host_auto_start = previous_auto_start
+            runtime._host_connection = previous_connection
+            runtime._host_snapshot = previous_snapshot
+
+    def test_blender_startup_does_not_start_ai_host(self):
+        runtime = self.addon.runtime
+        previous_auto_start = runtime._host_auto_start
+        previous_snapshot = runtime._host_snapshot
+        runtime._host_auto_start = False
+        runtime._host_snapshot = {"state": "STOPPED", "message": "Stopped"}
+        try:
+            with patch.object(
+                runtime,
+                "runtime_ready",
+                return_value=True,
+            ), patch.object(
+                runtime,
+                "_discover_live_host",
+                return_value=(None, None),
+            ), patch.object(
+                runtime,
+                "_close_finished_launched_host",
+            ), patch.object(runtime, "start_host") as start_host:
+                interval = runtime._service_host_timer()
+
+            start_host.assert_not_called()
+            self.assertEqual(interval, 2.0)
+            self.assertEqual(runtime.host_status()["state"], "STOPPED")
+        finally:
+            runtime._host_auto_start = previous_auto_start
+            runtime._host_snapshot = previous_snapshot
+
+    def test_models_panel_is_collapsed_and_assigns_each_download_target(self):
         draw_source = inspect.getsource(
-            self.addon.RuntimeModelsPanel.draw
+            self.addon.ModelsPanel.draw
         )
         model_source = inspect.getsource(
-            self.addon.RuntimeModelsPanel._draw_model
+            self.addon.ModelsPanel._draw_model
         )
 
         self.assertIn("for model in MODELS.values()", draw_source)
@@ -951,8 +1117,16 @@ class AddonTest(unittest.TestCase):
         self.assertNotIn("Isolated Runtime ready", draw_source)
         self.assertNotIn("License:", model_source)
         self.assertEqual(
-            self.addon.RuntimeModelsPanel.bl_label,
+            self.addon.ModelsPanel.bl_label,
-            "Setup",
+            "Models",
+        )
+        self.assertEqual(
+            self.addon.ModelsPanel.bl_parent_id,
+            self.addon.HostPanel.bl_idname,
+        )
+        self.assertEqual(
+            self.addon.ModelsPanel.bl_options,
+            {"DEFAULT_CLOSED"},
         )
 
     def test_ben2_debug_job_uses_its_image_and_model(self):
@@ -987,15 +1161,17 @@ class AddonTest(unittest.TestCase):
                 return_value=True,
             ) as model_ready, patch.object(
                 self.addon.runtime,
-                "remove_background_command",
+                "remove_background_request",
-                return_value=["worker"],
+                return_value=self.addon.runtime.HostJobRequest(
+                    {"job_id": "test"}
+                ),
             ) as remove_background:
                 command, prepared_paths = operator._prepare_command(
                     None,
                     settings,
                 )
 
-        self.assertEqual(command, ["worker"])
+        self.assertIsInstance(command, self.addon.runtime.HostJobRequest)
         self.assertEqual(prepared_paths, paths)
         model_ready.assert_called_once_with("BEN2_BASE")
         self.assertEqual(remove_background.call_args.args[1], input_path)
@@ -1030,15 +1206,17 @@ class AddonTest(unittest.TestCase):
                 return_value=True,
             ), patch.object(
                 self.addon.runtime,
-                "remove_background_command",
+                "remove_background_request",
-                return_value=["worker"],
+                return_value=self.addon.runtime.HostJobRequest(
+                    {"job_id": "test"}
+                ),
             ) as remove_background:
                 command, prepared_paths = operator._prepare_command(
                     None,
                     settings,
                 )
 
-        self.assertEqual(command, ["worker"])
+        self.assertIsInstance(command, self.addon.runtime.HostJobRequest)
         self.assertEqual(prepared_paths, paths)
         self.assertEqual(remove_background.call_args.args[1], input_path)
 

tests/test_ai_host.py

@@ -0,0 +1,223 @@
+import importlib.util
+import inspect
+import sys
+import tempfile
+import unittest
+from pathlib import Path
+from types import ModuleType, SimpleNamespace
+from unittest.mock import patch
+
+
+PROJECT_ROOT = Path(__file__).resolve().parents[1]
+
+
+def load_ai_host():
+    specification = importlib.util.spec_from_file_location(
+        "snapform_ai_host",
+        PROJECT_ROOT / "src" / "snapform" / "ai_host.py",
+    )
+    module = importlib.util.module_from_spec(specification)
+    specification.loader.exec_module(module)
+    return module
+
+
+class FakeCuda:
+    class OutOfMemoryError(RuntimeError):
+        pass
+
+    @staticmethod
+    def is_available():
+        return False
+
+    @staticmethod
+    def empty_cache():
+        return None
+
+
+class HostTest(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.host = load_ai_host()
+
+    def create_cache(self):
+        cache = self.host.ModelCache.__new__(self.host.ModelCache)
+        cache.torch = SimpleNamespace(cuda=FakeCuda())
+        cache.ben2_model = None
+        cache.ben2_key = None
+        cache.da3_model = None
+        cache.da3_key = None
+        return cache
+
+    def test_ben2_model_is_reused_for_matching_runtime_key(self):
+        loads = []
+
+        class FakeModel:
+            @classmethod
+            def from_pretrained(cls, model_dir):
+                loads.append(model_dir)
+                return cls()
+
+            def to(self, device):
+                self.device = device
+                return self
+
+            def eval(self):
+                return None
+
+        ben2 = ModuleType("ben2")
+        ben2.BEN_Base = FakeModel
+        cache = self.create_cache()
+        device = SimpleNamespace(type="cuda")
+        with patch.dict(sys.modules, {"ben2": ben2}):
+            first, _load_ms = cache.get_ben2("model", device)
+            second, cached_load_ms = cache.get_ben2("model", device)
+
+        self.assertIs(first, second)
+        self.assertEqual(loads, ["model"])
+        self.assertEqual(cache.status()["ben2"], "model")
+        self.assertEqual(cached_load_ms, 0.0)
+
+    def test_da3_model_is_reused_for_matching_runtime_key(self):
+        loads = []
+
+        class FakeModel:
+            @classmethod
+            def from_pretrained(cls, model_dir):
+                loads.append(model_dir)
+                return cls()
+
+            def to(self, device):
+                self.device = device
+                return self
+
+            def eval(self):
+                return None
+
+        package = ModuleType("depth_anything_3")
+        api = ModuleType("depth_anything_3.api")
+        api.DepthAnything3 = FakeModel
+        cache = self.create_cache()
+        device = SimpleNamespace(type="cuda")
+        with patch.dict(
+            sys.modules,
+            {
+                "depth_anything_3": package,
+                "depth_anything_3.api": api,
+            },
+        ):
+            first = cache.get_da3("model", device)
+            second = cache.get_da3("model", device)
+
+        self.assertIs(first, second)
+        self.assertEqual(loads, ["model"])
+        self.assertEqual(cache.status()["da3"], "da3_model")
+
+    def test_http_host_has_health_jobs_auth_and_single_task_guard(self):
+        source = inspect.getsource(self.host.create_app)
+
+        self.assertIn('@app.get("/health")', source)
+        self.assertIn('@app.post("/jobs", status_code=202)', source)
+        self.assertIn("x_snapform_token", source)
+        self.assertIn("hmac.compare_digest", source)
+        self.assertIn("parameters[\"status\"]", source)
+        self.assertIn('@app.post("/clients", status_code=201)', source)
+        self.assertIn('@app.post("/clients/{client_id}/heartbeat")', source)
+        self.assertIn('@app.delete("/jobs/{job_id}", status_code=202)', source)
+        self.assertIn('@app.post("/models/clear")', source)
+        self.assertIn('@app.post("/shutdown", status_code=202)', source)
+        self.assertNotIn("os._exit", source)
+
+    def test_host_tracks_clients_and_rejects_shared_shutdown(self):
+        fake_models = SimpleNamespace(
+            status=lambda: {"ben2": "", "da3": ""},
+            close=lambda: None,
+        )
+        with patch.object(self.host, "ModelCache", return_value=fake_models):
+            state = self.host.HostState("token", "instance")
+
+        state.register_client("first", "Blender A")
+        state.register_client("second", "Blender B")
+
+        self.assertEqual(state.snapshot()["clients"], 2)
+        self.assertEqual(state.request_shutdown("first"), "shared")
+        self.assertTrue(
+            state.reserve("job", "first", "reconstruct", "status.json")
+        )
+        self.assertFalse(
+            state.reserve("other", "second", "reconstruct", "other.json")
+        )
+
+    def test_cancelling_job_keeps_host_and_model_cache_alive(self):
+        fake_models = SimpleNamespace(
+            status=lambda: {"ben2": "BEN2_BASE", "da3": ""},
+            close=lambda: None,
+        )
+        with patch.object(self.host, "ModelCache", return_value=fake_models):
+            state = self.host.HostState("token", "instance")
+        state.register_client("client", "Blender")
+        with tempfile.TemporaryDirectory() as directory:
+            status_path = Path(directory) / "status.json"
+            state.reserve(
+                "job",
+                "client",
+                "remove-background",
+                str(status_path),
+            )
+
+            active = state.cancel_job("job", "client")
+            self.assertTrue(active["cancel_requested"])
+            self.assertTrue(state.is_job_cancelled("job"))
+            state.run_job(
+                {
+                    "job_id": "job",
+                    "command": "remove-background",
+                    "parameters": {"status": str(status_path)},
+                }
+            )
+            status = self.host.json.loads(
+                status_path.read_text(encoding="utf-8")
+            )
+
+        self.assertEqual(status["state"], "cancelled")
+        self.assertFalse(state.shutdown_event.is_set())
+        self.assertEqual(state.snapshot()["cached_models"]["ben2"], "BEN2_BASE")
+
+    def test_expired_clients_are_removed(self):
+        fake_models = SimpleNamespace(
+            status=lambda: {"ben2": "", "da3": ""},
+            close=lambda: None,
+        )
+        with patch.object(self.host, "ModelCache", return_value=fake_models):
+            state = self.host.HostState("token", "instance")
+        state.register_client("client", "Blender")
+        state.clients["client"]["last_seen"] -= (
+            self.host.CLIENT_TIMEOUT_SECONDS + 1.0
+        )
+
+        state.expire_clients()
+
+        self.assertEqual(state.snapshot()["clients"], 0)
+
+    def test_discovery_is_only_removed_by_matching_host(self):
+        with tempfile.TemporaryDirectory() as directory:
+            path = Path(directory) / "host.json"
+            args = SimpleNamespace(
+                discovery=str(path),
+                port=51234,
+                token="token",
+                instance_id="owner",
+            )
+            with patch.object(self.host.os, "getpid", return_value=42):
+                self.host.write_discovery(args)
+
+            payload = self.host.json.loads(path.read_text(encoding="utf-8"))
+            self.assertEqual(payload["port"], 51234)
+            self.assertEqual(payload["pid"], 42)
+            self.host.remove_discovery(path, "other")
+            self.assertTrue(path.is_file())
+            self.host.remove_discovery(path, "owner")
+            self.assertFalse(path.exists())
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_bootstrap_runtime.py

@@ -29,6 +29,7 @@ class BootstrapRuntimeTest(unittest.TestCase):
             args = SimpleNamespace(
                 uv="uv",
                 runtime_dir=str(runtime_dir),
+                models_dir=str(root / "models"),
                 backend="cu128",
                 da3_commit="test-commit",
                 ben2_commit="test-ben2-commit",
@@ -67,13 +68,48 @@ class BootstrapRuntimeTest(unittest.TestCase):
         self.assertTrue(
             any("test-ben2-commit" in str(value) for value in ben2_command)
         )
-        check_command = commands[-1]
+        host_command = next(
+            command
+            for command in commands
+            if f"fastapi=={bootstrap.FASTAPI_VERSION}" in command
+        )
+        self.assertIn(f"uvicorn=={bootstrap.UVICORN_VERSION}", host_command)
+        check_command = next(
+            command
+            for command in commands
+            if any("import torch;" in str(value) for value in command)
+        )
         self.assertTrue(
             any(
                 "from ben2 import BEN_Base" in str(value)
                 for value in check_command
             )
         )
+        self.assertTrue(
+            any("import fastapi, uvicorn" in str(value) for value in check_command)
+        )
+        default_downloads = [
+            command
+            for command in commands
+            if any(
+                repository in command
+                for repository in (
+                    "depth-anything/DA3-SMALL",
+                    "depth-anything/DA3-BASE",
+                    "PramaLLC/BEN2",
+                )
+            )
+        ]
+        self.assertEqual(len(default_downloads), 3)
+        ben2_download = next(
+            command
+            for command in default_downloads
+            if "PramaLLC/BEN2" in command
+        )
+        self.assertEqual(
+            ben2_download[-1],
+            '["config.json", "model.safetensors"]',
+        )
 
 
 if __name__ == "__main__":

tests/test_pack.py

@@ -38,5 +38,6 @@ class PackTest(unittest.TestCase):
         self.assertIn("image_depth.py", names)
         self.assertNotIn("snapform_depth_plane.py", names)
         self.assertIn("worker.py", names)
+        self.assertIn("ai_host.py", names)
         self.assertNotIn("importer.py", names)
         self.assertNotIn("snapform/__init__.py", names)

tests/test_worker.py

@@ -47,6 +47,20 @@ class WorkerTest(unittest.TestCase):
             source.index("import torch"),
         )
 
+    def test_ai_tasks_support_cooperative_cancellation(self):
+        with self.assertRaises(self.worker.TaskCancelled):
+            self.worker.check_cancelled(lambda: True)
+
+        self.worker.check_cancelled(lambda: False)
+        self.assertIn(
+            "cancel_check=cancel_check",
+            inspect.getsource(self.worker.reconstruct),
+        )
+        self.assertIn(
+            "cancel_check=cancel_check",
+            inspect.getsource(self.worker.remove_background),
+        )
+
     def test_optional_background_removal_runs_before_da3_model_load(self):
         source = inspect.getsource(self.worker.reconstruct)