Change8

v0.21.0

Breaking Changes
📦 comfyuiView on GitHub →
3 breaking13 features🐛 18 fixes🔧 17 symbols

Summary

This release introduces significant performance enhancements through Lora Async loading and dynamic VRAM management, alongside new features like Auto-regressive video generation and Gemma4 support. It also includes numerous bug fixes, particularly around media loading (pyav) and tensor handling, while removing IPEX support.

⚠️ Breaking Changes

  • The `save 3d model` node's filename prefix has changed from its previous value to `3d/ComfyUI` (CORE-106). Users relying on the old prefix for file location may need to update their workflows or expectations.
  • IPEX (Intel Extension for Pytorch) support has been removed (CORE-153). Users relying on IPEX must switch to standard PyTorch execution paths.
  • The Moonvalley API nodes have been removed from partner nodes (CORE-159).

Migration Steps

  1. If you were relying on the old filename prefix for the `save 3d model` node, update your workflows to expect the new prefix: `3d/ComfyUI` (CORE-106).
  2. Remove any configurations or dependencies related to IPEX (Intel Extension for Pytorch) as support has been removed.
  3. If using the Moonvalley API nodes, they must be removed or replaced.
  4. The `ref_image` input is now mandatory for the `ColorTransfer` node (CORE-140).

✨ New Features

  • Video loader node now supports reading audio and video simultaneously.
  • Added support for Gemma4 text generation (CORE-30).
  • Implemented Auto-regressive video generation (CORE-25).
  • Added high quality preview support for Flux2 latents.
  • Added support for OneTrainer ERNIE LoRA.
  • Added support for Topaz Astra 2 model via Partner Nodes.
  • Added Gpt 5.5 and 5.5-pro LLM models via Partner Nodes.
  • Added Luma UNI-1 model via Partner Nodes.
  • Added ByteDance: virtual portrait library for regular images via Partner Nodes.
  • Added support for dynamicVRAM with --cache-ram 2 (CORE-117).
  • Implemented block prefetch + Lora Async load + adoption in LTX for speedup (CORE-111).
  • Added generic --feature-flag CLI argument and --list-feature-flags registry.
  • Enabled triton comfy kitchen via cli-arg.

🐛 Bug Fixes

  • Fixed handling of un-inited meta-tensors in models, resolving a CPU TE crash (CORE-67).
  • Fixed SolidMask and MaskComposite device mismatch when using --gpu-only (CORE-98).
  • Fixed metadata rotation handling in pyav code.
  • Fixed pyav loading code to handle tRNS PNG.
  • Reduced video tiny VAE peak VRAM and decode time (CORE-127).
  • Lowered peak memory usage for 8 bit formats when using pyav.
  • Fixed SDPose resize issue (CORE-146).
  • Fixed ImageBlend and ImageCompositeMasked nodes to handle images with different channel counts (CORE-103).
  • Gracefully handled port-in-use error on server startup.
  • Fixed ColorTransfer node to make ref_image mandatory (CORE-140).
  • Fixed proper memory estimation for frame interpolation (CORE-150).
  • Fixed prefetch guard against no offload (CORE-149).
  • Made control-net load order deterministic (CORE-139).
  • Handled multi-compute of the same weight (CORE-153).
  • Fixed SplitImageToTileList and ImageMergeTileList to use tile_height for vertical stride minimum (CORE-152).
  • Fixed VAEDecodeAudio to be usable for LTX-2.x generated audio latents (CORE-157).
  • Fixed a typo in a1111 reference within extra_model_paths.yaml.
  • Fixed Content-Disposition header missing attachment prefix.

Affected Symbols