●Change8

v0.8.0

📦 comfyuiView on GitHub →
✨ 6 features🐛 10 fixes🔧 8 symbols

Summary

This release introduces support for sage attention 3, exposes several components in the public API, and includes numerous bug fixes, especially around model loading, memory management, and API nodes integration.

Migration Steps

  1. If you were relying on the old location of clip_preprocess, note that it has been moved to comfy.clip_model.
  2. If you encounter issues with fp4 operations, ensure your device supports them, or check CUDA version compatibility if using comfy_kitchen features.

✨ New Features

  • Added support for sage attention 3 in ComfyUI, enabled via a new CLI argument.
  • Exposed V3 Improvements, DynamicCombo, and Autogrow in the public API.
  • Added support for the LTXV 2 model.
  • Added LTXAVTextEncoderLoader node.
  • Added WAN2.6 ReferenceToVideo node in API Nodes.
  • Initial ops changes to use comfy_kitchen, including nvfp4 checkpoint support.

🐛 Bug Fixes

  • API Nodes: Preserved percent-encoding for signed URLs.
  • Ignored all frames except the first one for MPO format.
  • API Nodes: Tripo3D now passes face_limit parameter only when it differs from the default.
  • API Nodes: Added support for 720p resolution for Kling Omni nodes.
  • Fixed an issue where the upscale model wouldn't be moved to CPU.
  • Fixed a low VRAM issue with the ltxv2 text encoder.
  • Disabled ltxav previews.
  • Skipped fp4 matrix multiplication on devices that do not support it.
  • Disabled comfy kitchen CUDA if PyTorch CUDA version is less than 13.
  • Fixed an issue related to fp8 fast operations.

🔧 Affected Symbols

clip_preprocesscomfy.clip_modelmodel_managementTripo3D (API Nodes)Kling Omni nodes (API Nodes)ltxv2 text encoderltxav previewscomfy_kitchen