H3 Prompt & Size - SatoDive

Every ComfyUI workflow is mostly plumbing, and with MiniMax H3 the front of the graph is the worst of it: a text encoder, a size you have to compute in multiples of 32, reference images that need a different conditioning path than a plain text prompt, and a latent that isn't shaped like any normal image latent. H3 Prompt & Size does all of that and hands you two clean outputs - positive and latent.

Reach for it when you want to build your own sampler chain instead of using the pack's all-in-one nodes, or when you want the same conditioning feeding more than one branch. It's the front half of the graph, pre-assembled.

What it actually builds

Two things, and the second one is the interesting one.

The conditioning comes from ComfyUI's native H3 nodes. If any ref_image_* slot is connected it uses MiniMaxH3ReferenceToVideo (H3's reference/edit pathway, which is what makes <Picture N> work); with no refs it falls back to MiniMaxH3ImageToVideo. Either way it's called at length=5, the minimum H3 context, because for a still the conditioning doesn't depend on length - so you're not paying for frames you'll never sample.

The latent is a single-frame H3 still latent: one video frame with an even latent grid (the model's DiT patchifies 2×2, so the grid has to stay even) plus a small audio latent, wrapped together as a nested tensor. That's not an Image latent for a normal KSampler - it's an H3 latent, and it's the thing you'd otherwise be assembling by hand. H3 only supports batch size 1, so it's one latent, one image per pass.

The inputs you set

• prompt and the ref_image_1–ref_image_9 slots. The rule, straight from the tooltip: "N counts only connected slots, in slot order." Unused slots in the middle are invisible to numbering, so <Picture 1> is your first connected image.
• aspect + megapixels, or aspect = Custom with width/height (snapped to multiples of 32). Aspect Custom also ignores megapixels.
• ref_image_size - match scales refs down (never up) to the output area; max keeps them at a 2048px short edge for the best identity at a real speed and VRAM cost. The default here is match, which is the friendlier one while you're still hunting a composition.

There's no negative conditioning output and no CFG field, because H3 doesn't use one - the pack's samplers drive it with a single-conditioning guider. Don't go looking for a prompt-weight workaround; it isn't there.

Outputs

positive (CONDITIONING) goes to a guider or a sampler's positive input. latent (LATENT) is your starting image. Wire them into SamplerCustomAdvanced or into this pack's H3 Image Generation / H3 Draft Grid, which is what most people do.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/SatoDive/ComfyUI-H3-IMG-Gen-SatoDive

ComfyUI Manager will find it under ComfyUI-H3-IMG-Gen-SatoDive (MiniMax H3 Image Gen - SatoDive); restart after installing. Needs a ComfyUI build with the native H3 nodes, plus your H3 model, text encoder and VAE. No extra Python packages - the pack declares no dependencies at all.

Gotchas

Size maths is deliberately boring. Pick 16:9 at 3 MP and you get a multiple-of-32 pair that's near 3 MP, not exactly 3 MP - that's how the model wants it, and the exact-size fix happens downstream in the Simple or Final Size nodes.

ref_image_size is the main speed dial. Going max on a laptop GPU is the fastest way to make a workflow feel broken: the refs stay big through the text-and-VAE encode, which is exactly where small cards start swapping models in and out. Try match, then move back once the composition is locked.

One conditioning, one reuse. This node re-encodes prompt and refs every run. That's the right behaviour for a building block, but if you're re-rolling seeds in a hand-built graph, the pack's Simple node - which caches the last encode - will feel dramatically faster. That difference is the node, not your GPU.