ComfyUI v0.33.1: MiniMax Music 3 Native Support and CUDA Graphs

Native MiniMax Music 3 Support

The headline feature of v0.33.1 is native support for MiniMax Music 3, the open-weight music generation model. The built-in implementation generates full songs from lyrics and a sound description, including intro/verse/chorus/bridge structure with consistent vocal identity, up to 5 minutes long at 32kHz 16-bit stereo output.

The release notes also include follow-up fixes specific to the new model:

• Early qkv vs q,k,v detection for the MiniMax implementation
• Fixed MiniMax Music not working on non-dynamic VRAM systems

Core CUDA Graphs Support

v0.33.1 introduces core support for CUDA Graphs as part of the MiniMax Music 3 implementation. CUDA Graphs capture a sequence of kernel launches and replay them with minimal CPU overhead, which reduces per-step launch cost for repeated graph structures.

Partner Node Updates

• MiniMax: added ContextIR and Regenerate To 2K partner nodes
• Bria: added GenFill, Eraser, Expand and Increase Resolution partner nodes

Anima and Other Changes

• Anima tunes with extra blocks are now supported, extending the Anima family of image models
• Fixed KSamplerAdvanced with add_noise disabled on nested latents
• Fixed float64 device in the LTX diffusion decoder
• Dynamic VRAM is no longer disabled on WSL
• comfy-kitchen package updated to 0.2.31
• Workflow templates updated to v0.11.41

Getting v0.33.1

Update ComfyUI through the Manager or your launcher of choice. MiniMax Music 3 workflows require this version (or Comfy Cloud), since new model support ships tied to a version bump.