MiniMax-H3-HardMode-Workflow

MiniMax-H3 Hard Mode — reference to video

Identity from reference images, video and audio instead of a start frame. Up to 9 images, 3 videos, 3 soundtracks and 3 standalone audio clips.

This is the ref2va counterpart to the multishot workflows. Where those chain shots from a start frame, this one carries a person, a place or a voice from reference material you supply.

Also on Civitai.

v1.1 — chained shots, one master, with audio

By request: the Multishot release's chaining, driven by your references. H3_HardMode_Chained.json takes one script — one prompt per shot, --- between them. Shot 1 renders in Hard Mode from your reference images (and optional voice clip); every later shot continues image-to-video from the previous shot's last frame, so the identity cast in shot 1 carries down the chain in pixels; the segments join in-graph into one master with audio. Your <Picture N> binding lines go in their own box and attach to shot 1 automatically. Needs the node pack at v1.5+. At least 2 prompts; for a single shot use H3_HardMode_R2V.json.

What is here

You need the node pack — and nothing else custom

Every node in the graph is either ComfyUI core or from ComfyUI-H3-Multishot: H3ModelLoaderAny, H3ClipLoaderAny, H3ConditionStrength, H3FreeTextEncoder, H3ReferenceAudio. No other custom packs.

Install it, then run python apply_gguf_arch_patch.py from its folder and restart. Without that patch, loading an H3 GGUF fails with ValueError: Unexpected architecture type in GGUF file: 'minimax_h3' — ComfyUI-GGUF keeps a hardcoded architecture list and this one is not on it.

You also need a ref2va DiT (curve form or original), the text encoder with its mmproj sidecar, and the VAEs. fl2va files will not work in this mode.

Load the encoder with H3ClipLoaderAny at type = minimax, not with "CLIPLoader (GGUF)" — that node has no minimax type, falls back silently, and dies inside a tokenizer with an unrelated-looking JSON error.

Naming your references in the prompt

The part that is not guessable. Labels are 1-based, input slots are 0-based, so ref_image_0 is <Picture 1>:

<Picture 1> is the woman. <Picture 2> is the room she is standing in.
She walks to the window in the style of <Video 1>.

Ordering is fixed by the node however you wire it: all images, then videos, then standalone audios. A reference video with a soundtrack consumes an <Audio> ordinal before your standalone clips — one ref video with sound plus one ref_audio_0 makes your voice clip <Audio 2>, not <Audio 1>. Get that wrong and the wrong voice binds, with no error.

References and keyframes cannot be combined

A hard limit in ComfyUI core, not a choice made here. model_base.py writes cond_video_latents for keyframes, then assigns it again for references, so keyframe latents are discarded while the keyframe layout rows survive. The packed sequence desyncs and the sampler dies on a shape mismatch. Use the keyframes graph when you need a specific opening or closing frame.

Settings, measured

• Leave the eviction node ON. It frees the ~16.5 GB text encoder before the DiT loads. Images only: 3.4 min with it against 4.5 without, and the output is bit-identical — every pairing scored 0.00 mean pixel distance. With a reference video it is 6.5 min against a 46-minute run that was killed. If you ever disable it use Bypass, never Mute: a muted node produces nothing and the graph stops validating.
• ref_image_size = match, not max. Same subject and seed: max cost 24% more time and rendered a facial scar as a vague smudge on the wrong eyebrow, with less fine detail.
• Prefer more references over bigger ones. A third reference image cost nothing measurable — 2.82 min against 2.87.
• Reference audio must be STEREO. The audio VAE encodes [B, 2, L]; a mono clip produces half the rows it reserved and dies deep in the model with no useful error. The stereo-guard node handles it. Audio costs ~14% wall clock and trims motion ~14–27%.
• Reference video: minimum 5 frames, truncated to the generation length then trimmed to n % 17 == 5, shown to the encoder at 2 fps.

If a render crawls

Check power draw, not utilisation. A card thrashing weights between system RAM and VRAM still reports ~98% while pulling a fraction of its rated watts. Sample for 20–30 s though: the encoder phase at the start of every render is legitimately quiet, around 100 W, and one unlucky reading there looks exactly like thrashing.

Licence

Workflows MIT. Weights under MiniMax's licence. The reference images are synthetic, made for this release, and free to use.

Support

Everything I publish is free and stays free.

• Ko-fi
• GitHub Sponsors
• Liberapay