Lazy MiniMax All-in-One

MiniMax H3 is the reference-to-video model that went open-weight in mid-2026, and its big selling point is that one model family does text-to-video, image-to-video, first/last-frame video, and full reference-to-video from images, videos, or audio. The ComfyUI core nodes for that split into two conditioners: MiniMaxH3ImageToVideo for T2V/I2V/FL2V and MiniMaxH3ReferenceToVideo for R2V. Lazy MiniMax All-in-One wraps both behind one node and decides which path to run for you - which mode is active, which inputs matter, and what actually gets encoded.

If you've built an H3 graph the manual way, you know the pain. You swap conditioners when you switch modes, you re-wire frames, you keep two different samplers straight. This node collapses that into "connect everything once, pick a mode, hit queue."

How it works

Mode resolution is a short priority list, checked in order:

1. selector - a bare mode string (T2V, I2V, FL2V, R2V) or a tagged blob from the Lazy-subject-and-scene-automation node wins outright.
2. Otherwise, R2V if any reference inputs are connected.
3. Otherwise FL2V if both first_frame and last_frame are connected.
4. Otherwise I2V if first_frame is connected.
5. Otherwise T2V.

Once the mode is picked, the node hard-gates its sockets: T2V ignores all frames and references, I2V keeps only the first frame, R2V clears the first/last frames and expects references. You can leave everything wired and it'll use only what the mode needs - that's the "all-in-one" promise.

Under the hood it delegates to ComfyUI's native H3 nodes (MiniMaxH3ImageToVideo / MiniMaxH3ReferenceToVideo), with a vendored fallback for older ComfyUI. The pack needs ComfyUI 0.30+ for the native H3 integration; that's a hard requirement for this node to be useful.

The inputs that matter:

• clip / vae - required, from your H3 model loader.
• prompt - the text prompt (multiline).
• width (default 1344) / height (default 768) - resolution, snapped to multiples of 32.
• duration_seconds (default 5) - clip length; it's snapped to H3's 17k+5 frame grid at 24 fps, so 5s becomes 124 frames.
• first_frame / last_frame - IMAGE inputs for I2V / FL2V.
• ref_images, ref_videos, ref_video_audios, ref_audios - Autogrow reference sockets (R2V only) that expand one at a time as you connect.
• audio_vae - required for R2V, ignored otherwise. The tooltip says it straight: "Required for R2V (reference audio / video soundtrack encode)."
• selector - optional mode string that overrides socket guessing.
• ref_image_size - match or max, R2V only; forced to match when a tagged selector is present.

Outputs: positive (CONDITIONING), LATENT, and mode (STRING - handy for wiring into a switch so the rest of the graph can react to what actually ran).

Where people get burned

The two errors you'll actually hit are both explicit. R2V without audio_vae throws immediately - the model genuinely needs the audio VAE for reference encoding, so wire it even if your reference is a still image. And a zero or negative width/height raises a ValueError; the fix is in the node's own message - for T2V/R2V use a ResolutionSelector whose auto match includes t2v,r2v, and for I2V/FL2V let the size come from the first-frame loader.

Also: this node conditions and encodes - it doesn't load the H3 checkpoint or the audio VAE, and the pack ships no weights. Bring your own H3 model files. And expect the hardware ceiling: H3 is a heavy model, and the community consensus at launch was that you want real VRAM headroom, with reference-video being the most demanding mode.

Installing it

The node is part of vsaan212/Vsaan212-workflow-utilities (MIT). Install via ComfyUI Manager - search Vsaan212-workflow-utilities - or:

cd ComfyUI/custom_nodes
git clone https://github.com/vsaan212/Vsaan212-workflow-utilities

Restart ComfyUI and find it under vsaan212/minimax. No API keys, no cloud calls - the pack is fully local. The one real dependency is a current ComfyUI with native H3 support; update ComfyUI first if the node complains about missing H3 internals.