●Change8

v0.14.0

Breaking Changes
📦 comfyuiView on GitHub →
⚠ 2 breaking✨ 10 features🐛 12 fixes⚡ 1 deprecations🔧 11 symbols

Summary

This release introduces significant feature additions like the VideoSlice node, enhanced ControlNet support (Qwen 2512), and new API nodes for 3D and image background removal. It also includes numerous stability improvements, especially around dynamic VRAM management, training, and fixing issues with older PyTorch compatibility.

⚠️ Breaking Changes

  • Removed unsafe pickle loading code that was used for older PyTorch versions (older than 2.4). If you are using PyTorch < 2.4, you may encounter issues.
  • Removed workaround code that supported RMSNorm on old PyTorch versions.

Migration Steps

  1. If you rely on pickle loading for PyTorch versions older than 2.4, you must update your PyTorch installation.
  2. If you were using the "gpt-4o" model in API Nodes, you must select an alternative model.

✨ New Features

  • Added a new VideoSlice node.
  • Enabled Magnific Upscalers support in API Nodes.
  • Added support for generating attention masks for left padded text encoders.
  • Added left padding support to LTXAV text encoder.
  • Added working Qwen 2512 ControlNet (Fun ControlNet) support.
  • Introduced Node Replacement API.
  • Added ModelTo3DUV, 3DTextureEdit, and 3DParts nodes via API Nodes (Tencent).
  • Added Bria RMBG nodes via API Nodes.
  • Updated workflow templates to v0.8.42.
  • Implemented NAG (Normalized Attention Gradient) on all models based on Flux code.

🐛 Bug Fixes

  • Fixed vanilla-fp8 loaded lora quality issues when using dynamic_vram.
  • Fixed 3D models not showing in the outputs window.
  • API Nodes now retry on connection errors during polling instead of aborting.
  • Fixed Windows Aimdo crash and improved LLM performance related to dynamic_vram.
  • Fixed model dtype archival delay after override in SD operations.
  • Fixed an issue where ace15.py did not allow min_p sampling.
  • API Nodes now use a separate retry budget for 429 rate limit responses.
  • Fixed lora Extraction in offload conditions and dynamic_vram mode.
  • Fixed flux1 canny+depth lora crash by forcing load of flux img_in weight in MPDynamic.
  • Fixed anima preprocess text embeds not using the correct inference dtype.
  • Fixed some custom nodes issues.
  • Fixed training issues related to dynamic_vram.

Affected Symbols

VideoSliceModelPatcherDynamicace15.pyLTXAV text encoderQwen 2512 ControlNetModelTo3DUV3DTextureEdit3DPartsBria RMBG nodesllama (rope implementation)flux models (RMSNorm usage)

⚡ Deprecations

  • Removed model "gpt-4o" from api-nodes configuration.