ComfyUI News & Open-Source AI Releases

KoshiMazaki releases AKUSPACE v0.5, a spatial audio LoRA for LTX-2.5 that places generated voices and music in rooms, clubs, cathedrals or outdoors, with ComfyUI nodes included.

HR Endless Sampler is a ComfyUI custom node that renders unlimited-length MiniMax H3 videos by splitting inference into chunks, using Gemma to keep continuity.

MATLOWAI bakes the H3 turbo and Mystic LoRAs into a single INT8 ConvRot ComfyUI file: 4-step text/image-to-video and reference-to-video with audio, saving about 21 GB of VRAM.

minimax-h3-community-license-agreement

rzgar packages ByteDance's Bernini-Diffusers-v2 semantic planner and Wan2.2 renderer into ComfyUI custom nodes, with FP8 and NVFP4 weights and one workflow for six tasks.

apache-2.0

Fizgig v5.0.0 brings full fine-tuning of the MiniMax H3 33B and Krea 2 12.9B base models to consumer GPUs down to 16 GB VRAM, with a Checkpoint to LoRA exporter for ComfyUI.

LTX Ripple is a First Frame All Frames IC-LoRA for LTX-2.5: edit the first frame and the change propagates across the whole video while motion and timing stay intact.

A new ComfyUI PR adds an image-only LTXV 2.3 model path, with 13B Dev and Distilled checkpoints, CFG-free 8-step distilled sampling, and int8 convrot builds.

ComfyUI adds five Gemini Omni 1.1 Flash API nodes covering text-to-video, image-to-video, reference-to-video, scene extension, and video editing with 4K output.

NVIDIA releases a DMD-distilled 4-step preview of SANA-Video 2.0 5B, generating 720p 81-frame videos with BF16 CFG-1 inference and a four-stage sampling contract.

apache-2.0

Comfy announces it is the first official reseller of MiniMax H3 commercial-use licenses, covering video, audio and music models with full commercial rights and LoRA training.

FastVideo open-sources FastH3 Preview v1: a 4-step DMD2 distillation of MiniMax H3 with 90% sparse attention, up to 14x faster text-to-video-and-audio on Blackwell GPUs.

Studio 1939 is a golden-age animation style LoRA for MiniMax H3 with painterly and full-cel variants. Free weights, text-only character consistency, and two demo shorts.

Alibaba PAI ships PDD acceleration LoRAs for MiniMax H3: full audio-video generation in 8 or 4 sampler steps with no CFG, plus a native ComfyUI node pack.

apache-2.0

Alibaba's Wan 3.0 is live in ComfyUI through three API nodes: native 30-second single-pass video, 20-reference Omni Creation with @ syntax, and video editing by instruction.

Edit Anything v2 brings multi-instruction video editing to LTX-2.5: Replace, Remove, Add and Restyle in one prompt via a rank-128 IC-LoRA with a ready ComfyUI workflow.

MATLOWAI releases a rank-16 motion adapter for ComfyUI-MAINodes' H3 de-rope pass, retrained against real 240fps camera frames — plus Repair Plan/Splice and DyRoPE nodes.

ConvRot quantization runs SenseNova-U1.5-8B-MoT in ComfyUI on a 12 GB GPU: INT8 (17.6 GB) and hybrid W4A8 (13.8 GB) checkpoints plus an 8-step speed LoRA.

Apache-2.0

A ComfyUI node upscales long, high-resolution MiniMax H3 video with temporal chunking and spatial tiling, keeping VRAM bounded to one tile — plus an experimental LTX-2.5 variant.

A core-nodes-only ComfyUI workflow turns one scene photo into eight timed cinematic camera views in a single MiniMax H3 pass, then extracts frames and a 4x2 contact sheet.

Alibaba PAI releases a ControlNet-Union for MiniMax H3: one 6.8 GB checkpoint adds Canny, Depth, HED, MLSD and Pose video control plus inpainting via VideoX-Fun.

minimax-h3-community-license-agreement