v0.32.0
Breaking Changes📦 comfyuiView on GitHub →
⚠ 1 breaking✨ 9 features🐛 7 fixes🔧 10 symbols
Summary
This release enhances debugging of nested tensors, introduces new partner nodes for various models (Qwen, LTX, Grok), and improves API capabilities with new tag filters. It also includes several bug fixes for VRAM issues, model decoding, and dependency management.
⚠️ Breaking Changes
- Minimum officially supported PyTorch version is now 2.7. Older PyTorch versions may not be compatible.
Migration Steps
- Ensure PyTorch is updated to version 2.7 or later.
✨ New Features
- Added previewable_outputs_count to /api/jobs for Media Assets badge.
- Extended ER-SDE noise scaler by scaling h(t).
- Added Qwen-Image 3.0 Pro node.
- Added tags_all, tags_any, and tags_none tag filters to the assets list API.
- Implemented Comfy Kitchen attention.
- Added support for LTX 2.5.
- Added new nodes for LTX model version 2.5.
- Added Grok-Imagine-Image-2.0 model node.
- Mistral and Llama tokenizers no longer depend on the transformers library.
🐛 Bug Fixes
- Fixed upscale models breaking on non-dynamic VRAM with low VRAM.
- Fixed Create Layered Image discoverability and made its flags self-explanatory.
- Fixed VAEDecodeTiled crash on NestedTensor latents (MiniMax H3).
- Made the cu130 warning more visible.
- Fixed peak memory issue with H3.
- Fixed broken tiled audio decode.
- Fixed some clip vision regression.