v0.12.0
📦 comfyuiView on GitHub →
✨ 6 features🐛 11 fixes🔧 6 symbols
Summary
This release introduces significant performance improvements through adaptive model loading and KV cache support for llama models. It also includes numerous bug fixes across model compatibility, dynamic VRAM handling, and API nodes.
✨ New Features
- Added support for KV cache implementation when using llama models for text generation.
- Implemented adaptive model loading to reduce RAM usage, fix VRAM OOMs, and resolve Windows shared memory spilling issues.
- Added color type and Color to RGB Int node.
- Added API Nodes: RecraftCreateStyleNode, Vidu: Q3 model; Extend and MultiFrame nodes, HitPaw API nodes.
- Added basic support for the ace step 1.5 model.
- Enabled embeddings for some qwen 3 models.
🐛 Bug Fixes
- Fixed counting of non-dict items in outputs_count.
- Made empty hunyuan latent 1.0 work correctly with the 1.5 model.
- Fixed model not working with any resolution.
- Improved error message when a node type is missing.
- Fixed issues causing some custom nodes to break.
- Fixed dynamic_vram pinning when using model defined dtype.
- Fixed dynamic_vram to respect argument cast dtypes in non-comfy weights.
- Silenced pytorch buffer warning related to dynamic_vram.
- Fixed cast buffers issue when using intel offloading.
- Fixed Dynamic VRAM unloading.
- Fixed issue with parameters on the root model object.