toobusy MiniMax H3 Image Latent

MiniMax H3 is the 33B omni-modal video model that treats text, image, video and audio as one context - and its ComfyUI story has a gap. The official nodes generate video, but for single-image editing with the community T=1 image VAE there was no way to build the latent H3 demands, because H3 doesn't work on a plain image latent. It needs a one-frame video-plus-audio NestedTensor. toobusy MiniMax H3 Image Latent is a 25-line node that builds exactly that, so you can run H3's Ref2V (reference-to-image) path for still-image editing without patching ComfyUI core.

The "name is a lie" energy here is different from most nodes: the name is exactly what it does. It's not a VAE, not a sampler - just a latent constructor. It allocates a zeroed video tensor shaped [1, 24, 1, height/16, width/16] (24 channels, one frame, latent-space dimensions) and a zeroed audio tensor [1, 32, 2, 2], then wraps them in a NestedTensor. That's the entire mechanism, and it's all you need.

Inputs and outputs

• width - default 1344, multiples of 32.
• height - default 768, multiples of 32.

One output, LATENT, which you feed to SamplerCustomAdvanced alongside the H3 MODEL and CONDITIONING from the normal H3 loader nodes, then decode with the MiniMax H3 Image VAE. The README shows the whole chain:

width / height
-> toobusy MiniMax H3 Image Latent
-> SamplerCustomAdvanced
-> VAE Decode (MiniMax H3 Image VAE)

The layout is intentionally boring, and that's the point. The useful knowledge is in the constraints around it, not the node itself.

The constraints that bite

• Single frame, batch 1, period. This latent is for still-image editing. For actual video, use the official Empty MiniMax H3 AV Latent.
• A plain Empty Latent Image will not work. Its latent isn't an H3-style video+audio NestedTensor, and H3's conditioning pipeline rejects it. This is the failure mode people hit first.
• You need the model files. Nothing is bundled. You need an H3 Ref2V model and the MiniMax H3 Image VAE (community repack) in your diffusion_models / vae folders. Without the T=1 image VAE the whole single-image path is moot.
• Check the license before you build a pipeline on this. H3 ships under the MiniMax H3 Community License, which excludes the US, EU, UK and Korea from its applicable territory. That's a genuinely unusual geofence, and it applies to the weights you download, not just some API. Worth a minute of reading before you commit.

Installing it

This is part of the nicekriss/toobusy pack - install the pack, get this node:

cd ComfyUI/custom_nodes
git clone https://github.com/nicekriss/toobusy.git toobusy

or search "toobusy" in ComfyUI Manager, then restart ComfyUI and hard-refresh the browser. No extra Python dependencies for this node; it's pure torch against ComfyUI internals.

The changelog notes the author verified a six-edit character-sheet workflow on this path (docs/workflows/minimax_h3_single_image_6edit.json in the pack). If you're coming from the H3 video world and want cheap image-editing iterations without spinning up the full video pipeline, this is the most economical on-ramp the pack ships.