0.2.4

Pixal3D-ComfyUI 0.2.4

• Added subfolder support to filename_prefix.
• proj_1/p3d now exports to ComfyUI/output/proj_1/p3d_<timestamp>.glb.
• Added protection against absolute paths and directory traversal.
• Updated the node tooltip.

0.2.3

Pixal3D-ComfyUI 0.2.3 Patch Notes

Fixed

• Fixed native ComfyUI BiRefNet background removal failing with CPU/CUDA tensor mismatch during partial low-VRAM loading.
• Added a compatibility patch so BiRefNet relative-position attention tensors follow the active device and dtype during forward.

Improved

• Reused the DINOv3 image encoder across Pixal3D image-conditioning stages instead of building duplicate copies.
• Reused one loaded NAF model across the NAF stages instead of loading separate copies.
• Reduced duplicated helper-model memory during Pixal3D model loading.

0.2.2

Pixal3D-ComfyUI 0.2.2 Patch Notes

Fixed

• Fixed Windows/portable import errors caused by missing Pixal3D package __init__.py files.
• Fixed DINOv3 loading with stable transformers versions where encoder layers live under model.model.layer.
• Fixed DINOv3 CPU/CUDA tensor mismatch during image conditioning.
• Fixed full_gpu mode so it now forces a full Comfy model load instead of allowing partial CPU/GPU loading.
• Fixed non-square input stretching by automatically padding images to square before Pixal3D's square image encoder.
• Fixed older ComfyUI builds where comfy.ops.manual_cast.BatchNorm2d is missing.
• Lowered Pixal3D Export GLB minimum decimation_target to 5000; default remains 1000000.

Docs

• Added clear aspect-ratio padding order:
  • auto_remove: input -> RMBG/alpha crop -> pad to square -> RGB image sent to Pixal3D
  • keep_alpha: transparent input -> alpha crop -> pad to square -> RGB image sent to Pixal3D
  • none: input -> convert to RGB -> pad to square -> RGB image sent to Pixal3D
• Clarified that transparent images without RMBG should use background_mode=keep_alpha.
• Clarified that background_mode=none ignores alpha by design.
• Added low-poly export notes for decimation_target.

Not Included

• No new hole-filling feature was added in this release.

0.2.1

Pixal3D-ComfyUI v0.2.1

Added

• Added hybrid_low_vram mode to Pixal3D Model Loader.
• hybrid_low_vram combines Comfy/Aimdo-aware module construction with Pixal3D native stage-by-stage CPU/GPU offload.
• hybrid_low_vram now registers through Comfy model management, so it should appear in Aimdo/DynamicVRAM tensor visualizers.

Changed

• Kept native_low_vram as the older pure native staging path.
• native_low_vram still bypasses Comfy model management for maximum compatibility.
• Improved low-VRAM strict NAF handling by offloading the DINO backbone before the NAF forward spike.
• Offloads NAF after its forward pass in low-VRAM staged modes.
• Strict NAF now stays strict after a NAF failure instead of silently falling through on later runs.

Docs

• Updated README low-VRAM guidance to recommend hybrid_low_vram, with native_low_vram as fallback.
• Updated README_ZH with the same low-VRAM guidance.
• Updated troubleshooting docs for tight VRAM setups.

Version

• Bumped package version to 0.2.1.

Commit

• caa9298 Add hybrid low VRAM mode

0.2.0

Pixal3D-ComfyUI v0.2.0

Added

• Added a new rembg_image output to Pixal3D Image To 3D.
• Connect rembg_image to Preview Image to inspect the image Pixal3D used after background preprocessing.

Changed

• Improved Pixal3D Camera Control layout so the custom UI stays inside the node.
• Enabled resizing for Pixal3D Camera Control.
• Camera control resizing now keeps the node and preview canvas aspect ratio.
• Scrolling inside the camera UI no longer causes the widget to compress unexpectedly.

Docs

• Added a short rembg_image -> Preview Image note to README.md.
• Added the same note to README_ZH.md.

0.1.9

0.1.9

Fixed

• MoGe download path updated — Comfy-Org/MoGe restructured its repo from moge/ to geometry_estimation/. Downloads now fetch from the correct remote path and save to ComfyUI/models/geometry_estimation/
• Backward compatible — if you already have MoGe files in the old ComfyUI/models/moge/ location, the node still finds them. No re-download needed

Changed

• MoGe model local directory changed from ComfyUI/models/moge/ to ComfyUI/models/geometry_estimation/

0.1.8

Hotfix transformers compatibility
Cap transformers to <5.4 in requirements.txt and pyproject.toml to avoid current 5.x compatibility issues with the Pixal3D/RMBG stack.

0.1.7

Pixal3D-ComfyUI 0.1.7

Fixed

• Include the missing pixal3d/models package in the repository.
• Fix Pixal3DModelLoader failing with cannot import name 'models' from partially initialized module 'pixal3d'.
• Narrow .gitignore from models/ to /models/ so nested source folders like pixal3d/models are tracked.

Changed

• Cap transformers to <5 in requirements.txt and pyproject.toml to avoid current 5.x compatibility issues with the Pixal3D/RMBG stack.

0.1.5

0.1.5

Added

• Added Pixal3D Camera Control node for manual camera setup.
• Added a built frontend widget with Scene and POV views.
• Added manual_fov bundled camera output for a single cable into Pixal3D Image To 3D.
• Added low-VRAM camera-control example workflow:
  • example_workflows/pixal3d_low_vram_cam_control_example_workflow.json
  • example_workflows/pixal3d_low_vram_cam_control_example_workflow.png
• Added Linux / WSL CUDA requirements guide.

Changed

• Bumped package version to 0.1.5.
• Pixal3D Image To 3D now accepts optional manual_fov.
• When camera_mode=manual and manual_fov is connected, Camera Control overrides:
  • manual_camera_angle_x
  • manual_distance
  • mesh_scale
• When camera_mode=moge, connected Camera Control values are ignored.
• Updated Camera Control POV preview to match the manual FOV, distance, and scale sent to Pixal3D.
• Updated docs to recommend native_low_vram for low-VRAM workflows.

Docs

• Documented that Camera Control only works with manual camera mode.
• Added low-VRAM recipe:
  • vram_mode=native_low_vram
  • load_moge=false
  • load_rembg=false
  • use transparent PNG/WebP input
  • background_mode=keep_alpha
  • use Pixal3D Camera Control for camera setup
• Clarified Windows, Linux, and WSL CUDA wheel expectations.
• Clarified that FlashAttention remains a prerequisite.
• Added guidance that native low-VRAM can run around 4-8 GB VRAM for smaller workflows but needs roughly 40-50 GB system RAM.

Fixed

• Included web/ frontend registration for the Camera Control UI.
• Added node_modules/ to .gitignore.
• Improved tooltips for camera mode and manual_fov behavior.

0.1.3

Fixed

• Fixed native_low_vram CPU/GPU mismatch during DINO image conditioning.
• Replaced hardcoded .cuda() image placement with device-aware tensor movement.
• Improved native low-VRAM staging so Pixal3D helpers and pipeline modules move CPU/GPU more predictably.
• Added more aggressive cache cleanup after offload, OOM, MoGe use, RMBG use, and model destroy.

Changed

• Bumped package version from 0.1.2 to 0.1.3.
• --install-known-cuda now installs only Pixal3D CUDA extension wheels:
  • flex_gemm
  • cumesh
  • o_voxel
  • drtk
• FlashAttention 2 or 3 is now documented as a prerequisite, not something installed by the Pixal3D helper installer.
• Docs now clarify that plain natten==0.21.6 is only a baseline dependency. Strict NAF still requires natten.HAS_LIBNATTEN == True.
• Cleaned Windows wheel docs to separate:
  • FlashAttention prerequisite wheels
  • required Pixal3D CUDA extension wheels
  • optional renderer wheels
  • NATTEN/NAF limitations