FOSSWire
Search/
Local AI & Edge

ComfyUI v0.44.0 adds MiniMax-H3 support and expanded Partner Node integrations

Brief written by the FOSSWire AI editor from ComfyUI How we write

1 min read

Key Takeaway

ComfyUI v0.44.0 introduces extensive support for MiniMax-H3 models, new video upscaling targets, and various third-party 'Partner Nodes' for AI generation.

Summary

ComfyUI, the node-based GUI for stable diffusion and generative AI, has released version 0.44.0. This update focuses heavily on expanding the ecosystem through 'Partner Nodes,' which integrate specialized capabilities from various AI providers directly into the ComfyUI workflow.

Technical improvements include enhanced video handling, such as ProRes support and configurable chroma subsampling, alongside significant model support for the MiniMax-H3 architecture. The release also addresses low-level performance and stability, including fixes for VAE compute dtype crashes and optimizations for tensor comparisons in PyTorch.

Technical improvements in ComfyUI v0.44.0

  • Model Support: Added comprehensive support for MiniMax-H3, including the H3 latent upscaler, Sol-H3-VAE, and specific context window capabilities.
  • Video Processing: Introduced support for saving video in ProRes format and allows users to validate and select different chroma subsampling settings.
  • Partner Node Integrations: Expanded the ecosystem with new nodes for OpenAI (GPT-6.1 Terra and Sol), Ideogram (context window region editing), Topaz (Wonder 4 model), and VEED (Dubbing).
  • Compute & Performance: Implemented a new node to set VAE compute dtypes, fixed an issue where fp8_e4m3fn_fast would incorrectly fall back to bf16 during dynamic VRAM usage, and optimized tensor operations by skipping torch.equal calls when tensors are identical.
  • System Optimizations: Updated the development environment to use Python 3.12 for pylint and implemented a bypass for memfd_create to avoid Python overhead in certain file operations.

Why it matters

This release strengthens ComfyUI's position as a central hub for local and cloud-hybrid AI workflows by bridging the gap between specialized third-party models and a unified node interface. For developers, the improved video encoding and VAE dtype controls provide more granular professional-grade output options for generative media pipelines.

Original reference: github.com →