icrdr/AnyImage
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: Update documentation and improve depth processing logic; bump version to 0.11.1

Browse Source
This commit is contained in:
马南
2026-08-14 21:57:08 +08:00
parent 5255eb4e96
commit 102e1bfa9b
14 changed files with 149 additions and 34 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
src/snapform/blender_manifest.toml
View File

@@ -1,7 +1,7 @@
schema_version = "1.0.0"
id = "snapform"
version = "0.11.0"
version = "0.11.1"
name = "SnapForm"
tagline = "Build spatial structures from images, fast"
maintainer = "OmooLab"

21
src/snapform/da3_onnx.py
View File

@@ -15,6 +15,10 @@ except ImportError:
ONNX_PATH = "model.onnx"
PATCH_SIZE = 14
# DA3 cross-view attention is quadratic in (views x patches); the peak
# attention tensor alone can exhaust memory long before ONNX Runtime reports
# a generic allocation failure.
ATTENTION_BUDGET_BYTES = 4 * 1024 * 1024 * 1024
STRATEGY_CODES = {
"first": 0,
"middle": 1,
@@ -177,6 +181,14 @@ def camera_from_pose(pose_encoding, height, width):
return extrinsics, intrinsics
def estimate_attention_bytes(views, height, width):
"""Estimate the peak cross-view attention tensor size in bytes."""
patches = (height // PATCH_SIZE) * (width // PATCH_SIZE)
tokens = views * patches
# 12 attention heads in DA3 Base; float32 activations.
return tokens * tokens * 12 * 4
def infer(
session,
image_paths,
@@ -192,6 +204,15 @@ def infer(
process_res,
process_res_method,
)
_, views, _, height, width = model_input.shape
attention_bytes = estimate_attention_bytes(views, height, width)
if attention_bytes > ATTENTION_BUDGET_BYTES:
raise RuntimeError(
f"DA3 cannot process {views} frame(s) at {width}x{height}: "
"the cross-view attention would need about "
f"{attention_bytes // 2**30} GB. Reduce the frame count or "
"the Resolution Scale."
)
strategy_code = np.asarray(
[STRATEGY_CODES.get(ref_view_strategy, 2)],
dtype=np.int64,

72
src/snapform/image_depth.py
View File

@@ -1538,6 +1538,53 @@ def modifier_input_identifier(node_group, name):
raise RuntimeError(f"Geometry Nodes input is missing: {name}")
def modifier_input_slots(modifier):
"""Return the runtime input collection, or None for the legacy API."""
properties = getattr(modifier, "properties", None)
if properties is None:
return None
return getattr(properties, "inputs", None)
def set_modifier_input(modifier, identifier, value):
"""Assign a value to a Geometry Nodes modifier input socket."""
inputs = modifier_input_slots(modifier)
if inputs is not None:
# Blender 5.2+ exposes inputs as runtime RNA keyed by socket identifier.
getattr(inputs, identifier).value = value
else:
# Blender 4.x exposes inputs as modifier ID properties.
modifier[identifier] = value
def get_modifier_input(modifier, identifier):
"""Read a Geometry Nodes modifier input socket value."""
inputs = modifier_input_slots(modifier)
if inputs is not None:
return getattr(inputs, identifier).value
return modifier[identifier]
def modifier_input_driver_path(modifier, identifier):
"""Return the driver data path for a modifier input socket."""
if modifier_input_slots(modifier) is not None:
return (
f'modifiers["{modifier.name}"]'
f".properties.inputs.{identifier}.value"
)
return f'modifiers["{modifier.name}"]["{identifier}"]'
def drive_modifier_input(owner_object, modifier, identifier, expression="frame"):
"""Add a scripted driver to a modifier input so it follows the scene frame."""
driver = owner_object.driver_add(
modifier_input_driver_path(modifier, identifier)
)
driver.driver.type = "SCRIPTED"
driver.driver.expression = expression
return driver
def add_depth_plane_processing(
plane_object,
depth_image,
@@ -1565,12 +1612,10 @@ def add_depth_plane_processing(
("Cyclic", use_cyclic),
):
identifier = modifier_input_identifier(node_group, name)
modifier[identifier] = value
set_modifier_input(modifier, identifier, value)
if animate_frame:
frame_identifier = modifier_input_identifier(node_group, "Frame")
driver = modifier.driver_add(f'["{frame_identifier}"]')
driver.driver.type = "SCRIPTED"
driver.driver.expression = "frame"
drive_modifier_input(plane_object, modifier, frame_identifier)
plane_object.update_tag(refresh={"DATA"})
return modifier
@@ -1601,12 +1646,11 @@ def add_moge_depth_plane_processing(
("Frames", max(int(frame_duration or 1), 1)),
("Cyclic", use_cyclic),
):
modifier[modifier_input_identifier(node_group, name)] = value
identifier = modifier_input_identifier(node_group, name)
set_modifier_input(modifier, identifier, value)
if animate_frame:
frame_identifier = modifier_input_identifier(node_group, "Frame")
driver = modifier.driver_add(f'["{frame_identifier}"]')
driver.driver.type = "SCRIPTED"
driver.driver.expression = "frame"
drive_modifier_input(plane_object, modifier, frame_identifier)
plane_object.update_tag(refresh={"DATA"})
return modifier
@@ -1882,7 +1926,9 @@ class BuildDepthPlane(bpy.types.Operator):
assign_to_source=False,
)
color_image = None
if self.snapform_color_frames_dir:
# The foreground result already carries the color, so the original
# frame sequence is unused when background removal is available.
if self.snapform_color_frames_dir and foreground_image is None:
color_image = create_sequence_image(
f"{image_base_name(source_object.data)}.frame",
self.snapform_color_frames_dir,
@@ -2029,7 +2075,9 @@ class BuildMoge2DepthPlane(bpy.types.Operator):
assign_to_source=False,
)
color_image = None
if self.snapform_color_frames_dir:
# The foreground result already carries the color, so the original
# frame sequence is unused when background removal is available.
if self.snapform_color_frames_dir and foreground_image is None:
color_image = create_sequence_image(
f"{image_base_name(source_object.data)}.frame",
self.snapform_color_frames_dir,
@@ -2098,7 +2146,9 @@ class BuildOriginalDepthPlane(bpy.types.Operator):
assign_to_source=False,
)
color_image = None
if self.snapform_color_frames_dir:
# The foreground result already carries the color, so the original
# frame sequence is unused when background removal is available.
if self.snapform_color_frames_dir and foreground_image is None:
color_image = create_sequence_image(
f"{image_base_name(source_object.data)}.frame",
self.snapform_color_frames_dir,

7
src/snapform/image_mesh.py
View File

@@ -13,6 +13,7 @@ from .image_depth import (
process_resolution_scale_property,
remove_temporary_input,
scaled_process_resolution,
set_modifier_input,
video_total_frames,
)
@@ -417,7 +418,11 @@ def add_geometry_nodes_processing(
modifier.node_group = node_group
modifier.show_group_selector = False
identifier = modifier_input_identifier(node_group, "Resolution")
modifier[identifier] = default_remesh_resolution(vertices, faces)
set_modifier_input(
modifier,
identifier,
default_remesh_resolution(vertices, faces),
)
def create_matching_camera(

27
src/snapform/worker.py
View File

@@ -2016,14 +2016,25 @@ def reconstruct(args, model_cache=None, cancel_check=None):
0.25,
f"Estimating depth for {len(images)} frame(s)",
)
prediction = infer(
model,
[str(path) for path in images],
process_res=args.process_res,
process_res_method=args.process_res_method,
use_ray_pose=bool(args.use_ray_pose),
ref_view_strategy=args.ref_view_strategy,
)
try:
prediction = infer(
model,
[str(path) for path in images],
process_res=args.process_res,
process_res_method=args.process_res_method,
use_ray_pose=bool(args.use_ray_pose),
ref_view_strategy=args.ref_view_strategy,
)
except UnicodeDecodeError as error:
# DirectML reports failures with a locale-specific message that
# ONNX Runtime cannot always decode, masking the real error.
if provider_label(model) == "DirectML":
raise RuntimeError(
"DA3 inference failed on DirectML. Switch the device to "
"CPU in SnapForm Preferences and retry. "
f"Original error: {error}"
) from error
raise
check_cancelled(cancel_check)
write_status(args.status, 0.88, f"Exporting {args.export_format}")
if args.export_format == "mesh":