Painter MiniMax Ref To Video 6

MiniMax H3 is the 33B omni-modal model where picture and sound come out of the same pass - dialogue, effects and room tone generated together rather than glued on in a later audio stage. Great, and also means the workflow has a lot more wire-ends than a normal video graph: an image encoder, a video VAE, an audio VAE, reference frames, reference audio, and a text encoder that wants to see your photos.

Painter MiniMax Ref To Video 6 collapses the reference side of that into one node. Reference photos get uploaded straight into the node's own area (drag files in, multi-select, drop onto a slot to replace it). Reference video frames and audio still arrive on wires, because in practice those come out of other nodes - a loader, a frame extractor, a TTS pass.

The tags are the interface

The prompt is where you tell H3 what the references are for, using literal tags: <Picture 1>, <Video 1>, <Audio 1>. Reference material is registered as a token sequence before the text is encoded, so a prompt that never names <Picture 1> leaves that photo sitting in the context unused. The editor shows the same markup you typed, colour-highlighted, rather than rewriting it - which means you can box-select the finished prompt out of the node and reuse it elsewhere.

There's a rewriter built in. enable_prompt_optimize only shows/hides the opt_model, opt_api_url, opt_max_length and opt_template fields - the ✦ button in the editor corner triggers the rewrite either way. It talks to a local Ollama server (default http://127.0.0.1:11434, model qwen3.8-27b:latest, which you must pull, and needs vision). No cloud key, no API bill. The default opt_template rewrites your scribble into H3's six-section Ref2VA format: subject definitions, retention analysis, shot-by-shot description, soundscape.

Inputs that actually matter

Three encoders, and they're not interchangeable: clip (prompt + reference item list), vae (reference images and video frames into latents), audio_vae (reference audio, resampled with torchaudio if your clip isn't already at the audio VAE's rate). Then the shape of what you're about to generate:

• width / height - default 1376×768. The empty latent is built from these immediately: video is a 24-channel latent at 1/16 the resolution, with the audio latent running at 40 frames per second alongside it.
• length - frames at 24 fps, default 124 (≈5 s). The tooltip gives the trained range as roughly 124–362, i.e. 5 to 15 seconds. H3 only likes counts on the 17k + 5 grid and the node snaps you onto it internally, so 120 frames quietly becomes 124 - the step: 17 on the widget is a nudge, not the rule. Let Painter H3 Size Settings do the arithmetic.
• ref_max_size - default 1536. The long edge cap for every reference image and video: scaled down to fit, never up, then snapped to 32 px.
• ref_image_files - a JSON array of ComfyUI upload records. The frontend owns it; don't type in it. It's {filename, subfolder, type} pointing at real files in ComfyUI/input, which is why a workflow shared to another machine keeps the slots and loses the pictures.

The optional autogrow inputs take up to 3 each: ref_videos (frame batches at 24 fps, 2–15 s), ref_video_audios (matched by number - ref_video_audio_1 rides along with ref_video_1), and ref_audios for standalone clips you want cloned or lip-synced to. Reference videos under 5 frames raise an error, anything past your length gets truncated, the frame count is nudged down onto the 17k + 5 grid, and photos are loaded as RGB (EXIF rotation applied, alpha dropped), first frame only.

Outputs

positive → the sampler's positive. LATENT → the sampler's latent, already shaped for H3 audio-video, so don't bolt an EmptyLatent on behind it. prompt → a STRING pass-through of your text, handy for a save-text node.

v6 deliberately has no width/height/length outputs (v3 had them), so if something else needs those numbers, branch them off a size node rather than daisy-chaining this one.

Installing it

ComfyUI Manager → search PainterNodes, or clone the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/princepainter/ComfyUI-PainterNodes
cd ComfyUI-PainterNodes
pip install -r requirements.txt

That pulls soundfile>=0.12.1 and numpy>=1.21.0 - trivial. The node also imports torchaudio and PIL, which a normal ComfyUI environment already has. The real download is elsewhere: H3's DiT, video VAE, audio VAE and text encoder. And read the licence before you spend the disk - the MiniMax H3 Community License's Applicable Territory excludes the EU, UK, Republic of Korea and the United States, so in those regions you aren't licensed to run the local weights at all. That's a legal wall, not a technical one.

Where people get burned

The 33B checkpoint plus an audio branch plus multiple references is a lot of resident memory; no verified consumer VRAM floor was published at release and the weights land around 42.5 GB, so check your headroom before downloading. If you want ComfyUI to leave some of the card free while this runs, the pack ships Painter VRAM.