Support

Everything here is free and stays free — the format spec, the nodes, the workflows, the cartridges, the LoRAs. If it saved you a night of debugging (it contains several hundred of mine), tips keep the 5090 warm:

• ☕ Buy me a coffee on Ko-fi

• 💛 Sponsor on GitHub

• 🔁 Liberapay (recurring)

• ⚡ Or right here: the Civitai tip button on this page sends Buzz directly.

Identity from reference images, video and audio instead of a start frame.

Point it at photographs of a person, a place, a voice — up to 9 images, 3 videos, 3 soundtracks and 3 standalone audio clips — and the render carries them.

Three things you need besides this download. Both of the support questions this pack generated on day one were setup, not bugs, so they are first:

• The node pack — ComfyUI-H3-Multishot. Install via Git URL, restart, then hard-refresh the browser tab; the frontend caches node definitions and you will otherwise see red nodes. Then run python apply_gguf_arch_patch.py from its folder and restart again. Without that patch, loading any H3 GGUF fails with ValueError: Unexpected architecture type in GGUF file: 'minimax_h3' — ComfyUI-GGUF keeps a hardcoded list of architectures and this one is not on it.

• A ref2va DiT — curve form (smaller, needs ComfyUI 0.30.0+) or original form. fl2va files will not work in this mode.

• The text encoder and VAEs — Qwen3-VL encoder, and take the mmproj sidecar with it. Load it with the pack's H3ClipLoaderAny at type = minimax, not with "CLIPLoader (GGUF)" — that node has no minimax type, falls back silently, and dies deep inside a tokenizer with an unrelated-looking JSON error.

Naming your references in the prompt

This is the part that is not guessable. Reference blocks are labelled in the prompt text, and the labels are 1-based while the 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 regardless of how you wire it: all images, then videos, then standalone audios. Empty slots are skipped and the numbering compacts.

The ordinal trap. 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 it wrong and the wrong voice binds, with no error.

References and keyframes cannot be combined

Not a design choice here — a hard limit in ComfyUI's core. model_base.py writes cond_video_latents for keyframes, then assigns it again for references, so the keyframe latents are discarded while the keyframe layout rows survive. The packed sequence desyncs and the sampler dies on a shape mismatch. There is no "reference images plus start frame" mode, and asking for both crashes rather than degrading. Use the keyframes graph when you need a specific opening or closing frame.

Settings, measured rather than guessed

• 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 the difference between 6.5 min and a 46-minute run that was killed.

• Use ref_image_size = match, not max. Same subject, same seed: max cost 24% more time and rendered a facial scar as a vague smudge on the wrong eyebrow, with less fine detail. It does not buy fidelity.

• Prefer more references over bigger ones. Adding 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] and a mono clip produces half the rows it reserved, then dies deep in the model with no useful error. The stereo-guard node handles it; keep it wired. Audio costs about 14% wall clock and trims motion ~14–27%, which is the conditioning constraining the performance rather than stalling it.

• Reference video rules: minimum 5 frames, silently truncated to the generation length then trimmed to n % 17 == 5, and 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% utilisation while pulling a fraction of its rated watts. The tell in your log is:

loaded partially; 6423 MB usable, 5847 MB loaded, 19363 MB offloaded

Sample the wattage for 20–30 seconds before concluding anything. The encoder/conditioning phase at the start of every render is legitimately quiet — around 100 W at ~35% utilisation — and one unlucky reading there looks exactly like thrashing. If it climbs past 400 W once sampling starts, nothing is wrong.

What is in the download

• H3_HardMode_R2V.json — the reference-to-video graph, every node labelled

• workflows/ — the multishot AIO, long-form memory and keyframes graphs. These are also in the MiniMax-H3 Multishot pack you may already have; the copies here are relabelled so nothing shows a raw class name.

• reference_images/ — three views of one synthetic subject, so the graph works the moment you load it rather than pointing at ComfyUI's example.png

• HARD_MODE_GUIDE.html — the full guide, including the parts above

The reference images are synthetic, generated for this release, and free to use. Weights are under MiniMax's own licence; the workflows and node pack are MIT.