ComfyUI v0.45.0, released on 11 10, added the following to its "Open Source Model Support": text-to-image and editing templates for MiniMax-H3-Image, the WeVisOne-VL vision-language model, UltraTex for rebaking textures on 3D assets, RelayVSR video super-resolution, and video input for Qwen 3.x nodes. In terms of performance, Qwen 3.x merged parallel linear layer weights, SeedVR2 decoding switched to adaptive slicing to reduce VRAM usage, and Get Video Components now retains only a single copy of decoded frames. Additionally, issues such as Qwen FP8 text generation crashes and mesh retopology upon import were fixed. Version 0.45.2, released on 11 12, further fixed an issue where some workflows failed to load.
Why it matters In addition to MiniMax H3 Image, which was introduced a few days ago, this version adds usable nodes for local 3D and video understanding workflows, and reduces the VRAM required for SeedVR2 decoding.
License: Conditional commercial use · ComfyUI: Native support · VRAM: Not specified
Reasons
- Official changelog, with each open-source model support item linked to its corresponding merge request
Usage
- Update to v0.45.2 (which includes all content from v0.45.0) using the ComfyUI Manager or git pull, then find the corresponding workflow in the template library
Risks
- A single update involves multiple models; licenses are determined by each model's repository; RelayVSR's license has not yet been confirmed