ComfyUI News & Open-Source AI Releases

FastVideo ships FastH3 Trim, a 42-block pruned MiniMax H3 that runs 8-step video with audio in as little as 8 GB of GPU memory, with ComfyUI repacked files.

Kandinsky 6.0 open weights for ComfyUI: 29B Pro and 3B Lite generate 5-second video with synchronized 44 kHz audio, plus super resolution up to Full HD.

MIT

Prism from Fudan, Tencent Hunyuan and ZJU trains joint 2K video-audio models with dynamic sparse attention, 2.5x faster than full attention, with preview weights.

MIT

VNCCS 3.2 adds Qwen Image 2.1 and MiniMax H3 to its ComfyUI character pipeline, with transparent sprites, a rebuilt Character Creator and a 358-style library.

Fizgig v7 adds Anima and SDXL training, full fine-tuning for all six supported models down to an 8 GB card, and a complete driver system the community can extend.

A community LoRA for Qwen-Image 2.1 puts an outfit swap back on the original picture: framing, face and background hold while only the clothes change, with a ComfyUI workflow.

UltraTex from Jilin University and VAST textures 3D meshes at 2048x2048 with background token dropping and block-sparse attention, plus a ComfyUI node pack.

Scumble is a free, open-source desktop editor for AI inpainting: select an area, prompt a change and get a colour-matched layer back, rendered on your own ComfyUI or an API key.

GPL-3.0

The Veda team ships an official ComfyUI node that overrides MiniMax H3 attention with a distilled sparse predictor, skipping 90% of attention tiles for up to 3x faster video.

OmniChar ships an official ComfyUI node pack that loads a portable .char file and reuses one character identity across MiniMax H3, FLUX.2 and Krea 2 workflows.

ByteDance and Texas A&M release DMAD, a rank-128 LoRA that cuts MiniMax H3 to four steps for joint audio-video generation, with a Kijai ComfyUI conversion.

Comfy Org's Open Call Challenge invites the community to rebuild subscription-gated creative features as open-source workflows, with a $10,000 grand prize and RTX 5090 awards.

Comfy Agent plans, builds, runs and debugs ComfyUI workflows on the canvas itself, with reusable skills and up to 5 parallel chats. Live on Comfy Cloud now.

An experimental 2-in-1 release for MiniMax H3: a 2X video VAE plus a reference detail-enhancement node, with a tested ComfyUI workflow and a negative-result writeup.

MiniMax-H3-Community

ComfyUI-RMBG v3.2.0 adds a SAM 3.1 Multiplex node, an instant VRAM unload toggle, Apple Silicon GPU support, and cuts startup node scanning from 15s to under 0.1s.

ComfyUI Fantastic H3 Prompt Builder 1.8.0 adds masked clip editing with SAM 3.1, faster RefMod text encoding, and a reworked trim editor for MiniMax H3 prompts.

ComfyUI-Continuity 3.2 puts MiniMax H3, LTX 2.5 and five still families behind one node: managed casts, per-family RefMods, segment control and a selectable seam handoff.

AnyAngle Studio is a ComfyUI node that turns a photo, a GLB or a bundled mannequin into a 3D camera workbench and exports pose, depth and Canny guides for Qwen-Image 2.1.

UC San Diego and ByteDance release PDMD, a one-line change to DMD that distills MiniMax H3 to four steps, with a Kijai ComfyUI conversion.

Viggle's v0.3 distillation of Qwen-Image 2.1 adds a 9-step mode and ships official ComfyUI nodes, workflows and merged single-file weights for 6-step generation and editing.

Qwen Research License