What is this?
A community-maintained index of checkpoints, tools, and workflows for MiniMax H3, a video generation model that produces clips with native stereo audio from text, image, video, and audio inputs.
A community-maintained index of checkpoints, tools, and workflows for MiniMax H3, a video generation model that produces clips with native stereo audio from text, image, video, and audio inputs.Is this a tool, or just a list of links?It is a curated reference list (an awesome-list) that points to official and community checkpoints, quantizations, ComfyUI nodes, and workflows for MiniMax H3 — it does not run anything itself.
It is a curated reference list (an awesome-list) that points to official and community checkpoints, quantizations, ComfyUI nodes, and workflows for MiniMax H3 — it does not run anything itself.How is the list organized?It opens with a goal-based table (run locally, use audio, build in ComfyUI, write prompts, speed up, fine-tune, serve via API, run on Apple Silicon), then a VRAM table that maps GPU memory to a recommended checkpoint and text-encoder stack.
It opens with a goal-based table (run locally, use audio, build in ComfyUI, write prompts, speed up, fine-tune, serve via API, run on Apple Silicon), then a VRAM table that maps GPU memory to a recommended checkpoint and text-encoder stack.How much VRAM do I need to run MiniMax H3?Anywhere from 8 GB (DiffSynth-Studio NF4 path with heavy offloading) up to 24 GB+ for the standard pruned int8 ConvRot checkpoint; the README's VRAM table matches specific GPU tiers to checkpoint/text-encoder combinations.
Anywhere from 8 GB (DiffSynth-Studio NF4 path with heavy offloading) up to 24 GB+ for the standard pruned int8 ConvRot checkpoint; the README's VRAM table matches specific GPU tiers to checkpoint/text-encoder combinations.Are the quantized checkpoints official MiniMax releases?No. Only the original BF16 diffusers weights from MiniMaxAI are official; all the INT8/FP8/NVFP4/GGUF conversions in the list are community repackagings, so you should confirm compatibility with your runtime before downloading.
No. Only the original BF16 diffusers weights from MiniMaxAI are official; all the INT8/FP8/NVFP4/GGUF conversions in the list are community repackagings, so you should confirm compatibility with your runtime before downloading.Does this only work with NVIDIA GPUs, or can I use it on a Mac?Most stacks target NVIDIA GPUs from Turing (sm75) through Blackwell/RTX 50-series, but there is also a native Metal path for Apple Silicon via antirez/h3.c, with M3/M5 Max optimization still ongoing.
Most stacks target NVIDIA GPUs from Turing (sm75) through Blackwell/RTX 50-series, but there is also a native Metal path for Apple Silicon via antirez/h3.c, with M3/M5 Max optimization still ongoing.What's the difference between FL2VA and Ref2VA?FL2VA (first-and-last-frame mode) takes zero, one, or two input images and typically yields better raw quality; Ref2VA (omni-reference mode) accepts up to 9 images, 3 video clips, and 3 audio clips for richer conditioning but with slightly lower base quality that the Ref Patch partially offsets.
FL2VA (first-and-last-frame mode) takes zero, one, or two input images and typically yields better raw quality; Ref2VA (omni-reference mode) accepts up to 9 images, 3 video clips, and 3 audio clips for richer conditioning but with slightly lower base quality that the Ref Patch partially offsets.Can MiniMax H3 generate audio along with video?Yes, it generates video with native stereo audio directly from text, image, video, and audio inputs, and there is a dedicated comfyui-minimax-h3-audio-T8 node plus an audio VAE for this.
Yes, it generates video with native stereo audio directly from text, image, video, and audio inputs, and there is a dedicated comfyui-minimax-h3-audio-T8 node plus an audio VAE for this.How long and high-resolution are the generated clips?Clips can run up to 2K resolution and 15 seconds, and tools like ComfyUI-H3-Motion-Context can chain multiple blocks together for longer, multi-shot videos.
Clips can run up to 2K resolution and 15 seconds, and tools like ComfyUI-H3-Motion-Context can chain multiple blocks together for longer, multi-shot videos.Is there a one-command way to run MiniMax H3 locally?Yes, open-video-ai/open-video is described as 'Ollama for video models' — you install it, pull the model, and run it with a single command.
Yes, open-video-ai/open-video is described as 'Ollama for video models' — you install it, pull the model, and run it with a single command.What are similar projects?LTX-Video, CogVideo, LTX-2, FreeVideo, WanGP, stable-diffusion.cpp, hyperflow, Wan2.2
LTX-Video, CogVideo, LTX-2, FreeVideo, WanGP, stable-diffusion.cpp, hyperflow, Wan2.2Which topics is it in?blackwell, rtx, open-weights, image-to-video, model-serving, comfyui, audio-generation, gguf, text-to-video, metal, diffusion-models, lora
blackwell, rtx, open-weights, image-to-video, model-serving, comfyui, audio-generation, gguf, text-to-video, metal, diffusion-models, lora