RunningHub MiniMax H3 Video Gen (References)

Text prompts are only half of what MiniMax H3 does well. Feed it reference media - a face, a room, a camera move, a voice - and RHMiniMaxH3RefGen generates a new clip that actually honors them. This is the reference (Ref2VA) counterpart to RHMiniMaxH3VideoGen, and it's the node that makes H3 feel less like a text-to-video toy and more like a working edit tool: character stays the character, the set stays the set, and the audio can come from a reference clip too.

The limits, up front

H3's reference contract: up to 9 images, up to 3 videos, up to 3 audio references - 12 total - and audio can't be the only reference. The sequence is ordered images → videos → audio, and you reference them in the prompt by their per-type index: the first image is image_1, the first video video_1, and so on. A reference video with an audio track doubles as an audio reference automatically.

How it works

The node's inputs autogrow - images, videos, and audios are Autogrow ports, so you click to add more slots up to the limits. It builds the reference sequence, resolves the target shape from what you gave it (which is why aspect_ratio=auto is allowed here, unlike pure text), encodes the multimodal conditioning through the Qwen3-VL text encoder, and runs the same joint video+audio sampler as the other generation node. duration_seconds=0 derives the length from a single audio reference. Under the hood this replaces the whole legacy chain of Ref2VA reference nodes, target, and encode.

Inputs that matter

Beyond the media ports and prompt, the sampler bank matches RHMiniMaxH3VideoGen: seed (42, reproducible), steps (21 default - res_multistep at 21 points ≈ Euler-50 quality), video_shift (12) and audio_shift (3), sampler_mode (res_multistep recommended), and accel (off by default - profiles are calibrated at 50 steps, see the pack's sampling docs). Width/height of 0 derive from aspect_ratio. The ref_image_size option picks how reference images are sized into the sequence (match to the target, or max).

Optional overrides: h3_model, h3_text_encoder, h3_vae_bundle to hand it explicitly loaded components, sampler_config for the experimental knobs, and attention_backend (auto, sdpa, sage, or ck for Comfy Kitchen INT8).

Outputs

frames (IMAGE), audio (AUDIO), and av_latent (the joined latent, if you want to re-sample). Wire frames and audio to a video save node and you're done.

Install

Same story as every node in this pack - Manager (search "ComfyUI-RH-MiniMax-H3") 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 ~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 need ffmpeg/ffprobe on PATH here more than anywhere else in the pack - the node probes every video and audio reference you feed it.

Where people get burned

An audio-only reference set gets rejected - audio can't ride alone. A video_audio reference without an audio track throws rather than silently dropping the sound. And the same license reality check as the rest of H3: the Community License excludes the US, EU, UK, and South Korea, so check your region before building a pipeline on these weights.