RunningHub MiniMax H3 Video Gen (Text / Keyframes)

MiniMax H3 is the 33B omni-modal video model that generates picture and sound together - native stereo audio, not a music box bolted onto a silent clip. In ComfyUI, RHMiniMaxH3VideoGen is the front door. One node, three jobs, decided purely by what you plug into it: nothing plugged in is text-to-video-plus-audio (T2VA), first and/or last frame connected is keyframe generation (FL2VA), and a source_video is dubbing - the clip's picture becomes a clean condition and only the audio is regenerated.

No SGLang server, no Diffusers pipeline, no API key. The whole thing runs in your ComfyUI process.

How it works

The node is the modern all-in-one that replaced the old granular chain of target, text-encode, encode, and sampler nodes. Internally it's still that chain - it loads the FL2VA DiT, the Qwen3-VL text encoder, and the dual video/audio VAEs, resolves the output shape, builds conditioning, runs the joint video+audio sampler, and decodes - but you never see the plumbing. It even caches the static conditioning so re-runs skip re-encoding. The catch: this all-in-one version is only worth it if the models are actually there, because it auto-loads them from models/MiniMax-H3-INT8-CONVROT/ on first run.

The inputs that matter

• prompt - multiline. Write the visual and audio in one go; the model treats them as one context.
• aspect_ratio - 21:9 down to 9:16. auto only works when there's image/video material to reference; pure-text generation has to pick one.
• width / height - 0 means "derive from aspect ratio," and anything non-multiple-of-32 gets floored. Leave them at 0 unless you know why not.
• duration_seconds - 4–15. In keyframe mode it must be explicit; in dubbing mode 0 takes the source video's max usable slot.
• seed, steps, video_shift, audio_shift, sampler_mode, accel - the sampler bank. Defaults are sane: res_multistep at 21 sigma points gives roughly Euler-50 quality at about half the DiT passes, video_shift 12 / audio_shift 3 are the official flow shifts. accel defaults to off because the acceleration profiles are calibrated at 50 steps and overlap poorly with the 21-point path.

Optional first_frame and last_frame (keyframes, mutually exclusive with source_video) plus override inputs h3_model, h3_text_encoder, h3_vae_bundle if you want to hand it components from the loaders, and sampler_config if you're into the experimental knobs.

Outputs

Three, and they're what make this pack nice: frames (IMAGE), audio (AUDIO), and av_latent (the joined latent, if you want to re-sample instead of decode).

Install and models

Manager: search "ComfyUI-RH-MiniMax-H3" and hit install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/RH-RunningHub/ComfyUI-RH-MiniMax-H3.git
pip install -r ComfyUI-RH-MiniMax-H3/requirements.txt

Restart, then the real project: ~95 GiB of INT8 ConvRot weights into models/MiniMax-H3-INT8-CONVROT/:

cd /path/to/ComfyUI
hf download Gluttony10/MiniMax-H3-INT8-CONVROT --local-dir ./models/MiniMax-H3-INT8-CONVROT

You want ComfyUI 0.27+, a CUDA PyTorch build, and ffmpeg/ffprobe on PATH for the reference paths.

Where people get burned

aspect_ratio=auto on a pure-text run errors immediately - there's nothing to reference. Keyframe mode without an explicit duration also refuses. And before you point this at a 8GB card: Qwen3-VL 32B plus the DiT is a big resident footprint; the INT8 weights help, but you still need serious VRAM or patience with offload. One more thing to know up front: the MiniMax H3 Community License excludes the US, EU, UK, and South Korea - if you're in those regions you're not licensed to run these weights locally, even though the pack installs fine.