●Change8

v0.4.0

📦 comfyuiView on GitHub →
✨ 11 features🐛 33 fixes⚡ 1 deprecations🔧 14 symbols

Summary

This release introduces the Ovis image model, significant VRAM reductions via temporal rolling VAE, and extensive migration to the V3 Schema for several node groups. It also includes numerous bug fixes focused on memory management, model loading, and compatibility across various models like Kandinsky and Qwen.

Migration Steps

  1. Replace absolute imports with relative imports in comfy_api (if applicable to your custom code).
  2. Replace imports of deprecated V1 classes (if applicable to your custom code).

✨ New Features

  • Implement the Ovis image model.
  • Support ComfyUI-Manager for pip version.
  • Implement temporal rolling VAE (Major VRAM reductions in Hunyuan and Kandinsky).
  • Add MatchType, DynamicCombo, and Autogrow support to V3 Schema.
  • Add support for "@image" reference format in Kling Omni API nodes.
  • Convert nodes_load_3d.py to V3 schema.
  • Convert nodes_audio.py to V3 schema.
  • Add System Instruction support & Visual Enforcement for Banana Nodes in API Nodes.
  • Add support for seedance-1-0-pro-fast model in API Nodes.
  • Support "transformer." LoRA prefix for Z-Image.
  • Add chroma-radiance-x0 mode.

🐛 Bug Fixes

  • Fix CODEOWNERS formatting to have all on the same line.
  • Add check for the format arg type in VideoFromComponents.save_to function.
  • Hack to make zimage work in fp16.
  • Use flag based OOM fallback for attention.
  • Support Z Image alipai controlnets.
  • Fix issue with portable updater.
  • Reduce massive Lora reservations especially on Flux2.
  • Use empty list defaults for Schema.inputs/outputs/hidden to avoid None issues.
  • Convert nodes_audio.py to V3 schema.
  • Qwen Image Lora training fix.
  • Fix case where text encoders where running on the CPU instead of GPU.
  • Use look-ahead actuals for stream offload VRAM calculation (fixes unwanted TE full offload).
  • Revise HY1.5 VAE estimate.
  • Fix transformer_options getting cleared in Lumina model after z-image PR.
  • Make old scaled fp8 format use the new mixed quant ops system.
  • Fix regression when text encoder loaded directly on GPU.
  • Context windows fixes and features.
  • Fix some custom nodes.
  • Fix regression.
  • Convert nodes_freelunch.py to the V3 schema.
  • Convert nodes_mask.py to V3 schema.
  • Set OCL_SET_SVM_SIZE on AMD.
  • Fix Wrong EmptyAudio node input types.
  • Speed up lora compute and lower memory usage by doing it in fp16.
  • Fix on-load VRAM OOM.
  • Fix qwen scaled fp8 not working with kandinsky. Make basic t2i wf work.
  • Properly load the newbie diffusion model.
  • Retune lowVramPatch VRAM accounting.
  • Fix regression.
  • Fix potential issue.
  • Dequantization offload accounting (fixes Flux2 OOMs - incl TEs).
  • Fix for HunyuanVideo1.5 meanflow distil.
  • Fix nan issue when quantizing fp16 tensor.

🔧 Affected Symbols

VideoFromComponentszimageVideoFromComponents.save_toLumina modelKling Omni API nodesnodes_load_3d.pynodes_audio.pyBanana Nodesseedance-1-0-pro-fast modelnodes_freelunch.pynodes_mask.pyEmptyAudio nodenewbie diffusion modelcomfy_api

⚡ Deprecations

  • Pika API nodes are marked as deprecated.