Painter MiniMax Ref To Video 3

MiniMax H3 - the 33B omni-modal video model that went open-weights in August 2026 - is genuinely great at the job nobody else could quite do locally: hand it a couple of reference pictures plus a loosely written idea, and it returns video with synced audio, obeying instructions like "cut at 3.5 seconds." The community called its motion transfer "like Kling". Feeding it references from ComfyUI was the fiddly part, though - load every frame yourself, hand-build the conditioning.

Painter MiniMax Ref To Video 3 is one node that swallows the whole job: upload reference images inside the node, type your idea, and out come the CONDITIONING and LATENT that feed H3's sampler. It's the third revision, and the version history is basically the history of what annoyed people - v1/v2 did multi-reference conditioning with external inputs; v3 moves the images to an in-node upload area and adds a prompt optimizer.

How it works

Under the hood it's a conditioning-and-latent factory. It takes the three model pieces H3 splits into - clip, vae (the video VAE) and audio_vae. It builds an empty joint audio+video latent from your width/height/length, snapping the frame count to H3's temporal grid (frames such that n % 17 == 5) and deriving the matching audio-latent length at 40 fps. Reference images get scaled down so their longest edge fits ref_max_size (never upscaled), snapped to 32px and VAE-encoded; reference videos are encoded as frame sequences and sampled down to ~2 fps keyframes so the model's vision can actually read them; reference audio is resampled and encoded with the audio VAE. Then it tokenizes your prompt with all those references attached and stamps the minimax_refs into the conditioning. What leaves the right side feeds the sampler: positive → conditioning, LATENT → latent, then decode with VAEDecode plus VAEDecodeAudio.

This is one Painter node where the work is real and checkable in the source - it genuinely encodes refs and builds an AV latent. That matters: the author (princepainter, a prolific one-person node factory that wraps new video models within days) has drawn occasional "does this node actually do anything?" scrutiny from the community.

What you actually set

• prompt - the fun part. The custom editor speaks the shorthand that maps onto MiniMax's official H3 prompting skill: type @图片1 / @音频1 / @视频1 (rendering as <Picture 1> / <Audio 1> / <Video 1> tags) to reference an uploaded asset, wrap dialogue as 【台词内容】, write 切镜3.5 for a cut at 3.5 seconds. The UI labels are Chinese-first; the inserted tags are the English ones H3 expects.
• length (default 124) - frame count at 24 fps. 124 is about 5 seconds; the trained sweet spot is roughly 124–362.
• width / height (default 1376×768) and ref_max_size (default 1536, the cap for reference long edges) are set-and-forget.

The optimizer block is v3's other headline: a ✦ button on the prompt editor calls a local Ollama model to rewrite your rough idea into H3's six-section prompt format - and because the model can see your reference images, the rewrite actually describes the character you locked in. You need Ollama running locally, and opt_model should be the name of a vision-capable model you've actually pulled. Be warned: the default template is a long Chinese system prompt and the rewrite comes back in Chinese until you edit it - the enable_prompt_optimize toggle just shows and hides that block.

Outputs: positive and LATENT for the sampler, plus width / height / length / prompt - handy bookkeeping, and prompt shows you what the optimizer produced.

Installing it

In ComfyUI-Manager, search "PainterNodes" and install, or:

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

Then restart. The pack's own dependencies are just soundfile and numpy, but this node is written against ComfyUI's new declarative node API (io.ComfyNode with autogrow inputs), so you need a genuinely recent ComfyUI - if the node misbehaves, update ComfyUI before anything else. You also need the H3 weights from MiniMaxAI/MiniMax-H3 (~42.5 GB at full precision) in your workflow's loaders, and Ollama only if you want the optimizer.

Where people get burned

• The license. H3's Community License excludes the US, EU, UK and Korea from running the open weights locally. In those regions the hosted Hailuo API is the licensed path; no node fixes that.
• Reference videos need ≥ 5 frames (~0.2 s at 24 fps) or the node raises.
• Audio VAE unplugged → silent or failing output. It's a required input for a reason.
• Workflow version drift. This pack iterates weekly (v1 → v2 → v3 in a month). A saved workflow from v1/v2 expects external image inputs; v3 wants them uploaded inside. If links won't attach, that's why - re-wire, don't debug.