v0.9.0
📦 comfyuiView on GitHub →
✨ 9 features🐛 12 fixes🔧 3 symbols
Summary
This release introduces new nodes like JoinAudioChannels and image compare, improves model loading compatibility (including new LoRA formats), and fixes several issues related to attention, VAE encoding, and mixed precision operations.
Migration Steps
- If you rely on specific behavior related to sequential CI test jobs, note that this behavior has been reverted.
✨ New Features
- Added node: JoinAudioChannels.
- Added basic Asset support for models.
- Added image compare node.
- Added 'cancelled' status for interrupted jobs in /api/jobs endpoint.
- Added API Nodes for Vidu2.
- Enabled PyTorch Attention by default on gfx1200 hardware.
- Support ModelScope-Trainer/DiffSynth LoRA format for Z-Image models.
- Support the siglip 2 naflex model as a clip vision model.
- Refactored code to attempt lowering memory usage.
🐛 Bug Fixes
- Reverted change that forced sequential execution in CI test jobs.
- Improved detection for AMD torch compiled with efficient attention.
- Fixed import issue.
- Fixed CSP error in frontend when forcing offline mode.
- Fixed VAEEncodeForInpaint to correctly support WAN VAE tuple downscale_ratio.
- Fixed issue where Topaz Enhance API node downscaled the input image.
- Fixed saving of mixed ops weights.
- Fixed issue with t5 text encoder in fp4 precision.
- Fixed chroma fp8 te being incorrectly treated as fp16.
- Fixed API node issue requiring a unique name for uploading audio files.
- Made loras work correctly on nvfp4 models.
- Made database bulk_ops.py more widely compatible across sqlalchemy and python versions.