
ComfyUI v0.38.0 released: Full-chain optimization for Qwen-Image 2.1, with integration of new nodes like Yuan-3.5.
ComfyUI releases version 0.38.0, incorporating over 60 changes: This includes the compilation and memory optimization of Qwen-Image 2.1, improvements to the inference engine and asset subsystem, as well as the addition of collaborative node models from Tencent's Hunyuan Image 3.5, Byte Seedream 5.0 Flash, and GPT-6.
Related tags
Overview of Release
ComfyUI is an AI creative engine for visual professionals, centered around a modular node graph interface, capable of generating images, videos, and 3D models, and primarily targeting local consumer-grade hardware. It’s touted as the “most optimized inference engine for diffusion models globally.” On September 29, 2026, ComfyUI released version v0.38.0, which is another major iteration following v0.37.0, encompassing over 60 changes, with a focus on full optimization of Qwen-Image 2.1, inference engine and memory optimization, new model integration, and updates for partner nodes from Tencent, OpenAI, Anthropic, and ByteDance.
Key Highlights
1. Full Optimization of Qwen-Image 2.1
- Improved KV cache location logic for Qwen 2.1 to reduce memory pressure during long sequence inference (#16429).
- Enabled compilation for Qwen Image 2.1’s transformer blocks, improving inference speed (#16430).
- Supported the tiny VAE of Qwen-Image 2.1 to reduce overhead in the VAE decoding phase (#16552).
- Added support for Qwen-Image 2.1’s union fun controlnet (#16519).
- Implemented in-place clamping of Qwen Image 2.1’s fp16 activation values to prevent numerical overflow (#16608).
2. Inference Engine and Memory Optimization
- Allowed models to declare which attention implementation should be used for each block, introducing comfy_attention and AttentionTensorContainer across multiple models (#16419, #16595).
- Extended fast disk detection to all model loaders (#16425).
- Ported optimizations to the flux model family, and reduced memory usage for Lumina series models (#16488, #16515).
- Added support for NPU asynchronous weight offload streams (#16057).
- Optimized Seedvr2, adding support for w6a8 quantization format ([#165
More News
Want to see more quality content?
Browse all news