v0.15.0
Breaking Changes📦 comfyuiView on GitHub →
⚠ 2 breaking✨ 10 features🐛 9 fixes🔧 9 symbols
Summary
This release introduces basic native text generation support (Gemma3), adds several new nodes including GLSL and audio equalizer, and includes numerous fixes related to fp8 performance, Gemini image handling, and LTXAV compatibility.
⚠️ Breaking Changes
- LTXAV workflows might be temporarily broken due to LTX 2.0 migration; users should migrate LTXAV workflows to LTX 2.0.
- LTXAV embedding connectors were moved to diffusion model, which may break existing workflows relying on the old location.
Migration Steps
- Migrate LTXAV workflows to LTX 2.0 to ensure compatibility.
- Update nightly installation command for ROCm as specified in the release notes.
✨ New Features
- Added basic text generation support with native models, initially supporting Gemma3.
- Added GLSL shader node using PyOpenGL.
- Added gradient-slider display mode for FLOAT inputs.
- Added essential subgraph blueprints.
- Added support for pyopengl < 3.1.4 where the size parameter does not exist in GL context creation.
- Added SplitImageToTileList and ImageMergeTileList nodes.
- Added KlingAvatar node via API Nodes.
- Added ByteDance Seedream-5 model via API Nodes.
- Added ElevenLabs nodes via API Nodes.
- Added simple 3 band equalizer node for audio.
🐛 Bug Fixes
- Fixed performance of some fp8 dynamic_vram workflows by limiting return of requants.
- Fixed Gemini image MIME type handling by using glob matching.
- Fixed API Nodes to force Gemini to return uncompressed images.
- Fixed non-contiguous audio waveform crash when saving video.
- Fixed issue where subgraph files were not read with UTF-8 encoding.
- Fixed dtype issue in embeddings connector.
- Fixed issue loading fp8 ltxav checkpoints.
- Fixed crash when an operation is uninitialized.
- Swapped essentials_category assignment from CLIPTextEncode to PrimitiveStringM.