v0.11.0
Breaking Changes📦 comfyuiView on GitHub →
⚠ 2 breaking✨ 20 features🐛 14 fixes🔧 12 symbols
Summary
This release introduces significant feature additions, including support for zimage omni, Qwen 3 0.6B, and various API nodes, alongside numerous performance improvements and bug fixes, particularly around LTX2 and VAE memory usage.
⚠️ Breaking Changes
- Fix mistral 3 tokenizer code failing on latest transformers version and other breakage. Users relying on the old tokenizer implementation for Mistral 3 models may need to verify compatibility with the updated transformers version.
- The ByteDanceImageEditNode (seededit) node has been removed from API Nodes. Users should migrate workflows using this node to an alternative method.
Migration Steps
- If using the ByteDanceImageEditNode (seededit) in API Nodes, migrate workflows to use an alternative node.
- Verify compatibility if relying on the specific implementation of the mistral 3 tokenizer that was updated to fix breakage with the latest transformers version.
✨ New Features
- Support for zimage omni base model.
- Support for WaveSpeed nodes in API Nodes.
- Dynamic detection of chroma radiance patch size.
- Configuration support for Qwen 3 0.6B model.
- LyCORIS LoKr support added for Flux2.
- Vidu Reference node in API Nodes now allows passing up to 7 subjects.
- Added 'search_aliases' field to node schema.
- Support for the Anima model.
- Support for LTX2 Tiny VAE (taeltx_2).
- Support for Multi/InfiniteTalk.
- Support for ply & 3dgs formats in 3D nodes.
- Support loading flux 2 klein checkpoints saved with SaveCheckpoint.
- Propagation of Qwen-Image attention mask to image generation.
- Search aliases added to all nodes.
- Improved ResizeImageMaskNode UX with tooltips and search aliases (re-applied after revert).
- Support for ModelScope-Trainer/DiffSynth LoRA format for Flux.2 Klein models.
- Support for TencentHunyuan3D nodes in API Nodes.
- Added support for kwargs inputs to allow arbitrary inputs from the frontend.
- Added RoPE Position Adjustment for Small Grid IC-LoRA.
- Added Magnific nodes in API Nodes.
🐛 Bug Fixes
- Removed normalization of audio in LTX Mel spectrogram creation.
- Fixed edge case of EasyCache when conditionings change during a sampling run.
- Reduced warnings when loading embeddings_connector weights for LTX2.
- Reduced LTX2 VAE VRAM consumption.
- Fixed spatial inpainting issue in LTX2 refactor.
- Speed up and reduced VRAM usage for QWEN VAE and WAN.
- Made regular empty latent node work properly on flux 2 variants.
- Made empty latent node work with other models.
- Only enable fp16 on z image models that actually support it.
- Bypassed forward mode in Weight adapter system for Trainer.
- Added name to LoraLoaderModelOnly.
- Fixed Noise_EmptyNoise when using nested latents.
- Switched off feature cache for single frame in wan-vae.
- Updated amd portable for rocm 7.2.