RunningHub MiniMax H3 FL2VA Target (Legacy)

Every MiniMax H3 generation needs a target - the agreed-upon shape and length that conditioning, latent, and sampler all conform to. This node is how you set that for FL2VA (first/last-frame) generation: it takes your keyframes plus an aspect ratio and duration, and produces the MINIMAX_H3_TARGET that flows through the encode, latent, and sampler nodes. Think of it as the resolution picker on a standard sampler, upgraded to carry the whole geometry contract.

Inputs

• keyframes - required, the MINIMAX_H3_FL_KEYFRAMES from a first/last-frame condition node. The target can read intended composition off these.
• aspect_ratio - auto (recommended), 21:9, 16:9, 4:3, 1:1, 3:4, 9:16. auto derives the ratio from your keyframes, which is the right default for first/last-frame work - your input images are the composition, so forcing a ratio is usually asking for a fight.
• duration_seconds - 4.0 to 15.0, default 5. H3's range, straight from the model card: clips from 4 to 15 seconds. Longer isn't necessarily better - longer means more frames and a heavier sampler bill.

The optional width and height are the manual override: if both are 0, the aspect ratio governs; set them (multiples of 32) and you're choosing exact pixels instead. Handy when a video pipeline downstream wants a specific resolution.

Outputs:

• target - the MINIMAX_H3_TARGET, wired into the FL2VA Encode node and the Empty AV Latent.
• shape_info - a STRING describing the resolved shape. Mostly a debugging aid; hook it to a text display node if you want to see exactly what H3 computed for your frames.

What "auto" is doing

H3's target resolution logic scales around the keyframes and clamps to the model's sensible canvas, and shape_info will show you the verdict - width, height, frame count, and the latent sizes that flow out of it. If your keyframes are wildly different resolutions, auto will pick a compromise; that's the moment to set explicit width/height instead.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/RH-RunningHub/ComfyUI-RH-MiniMax-H3.git
pip install -r ComfyUI-RH-MiniMax-H3/requirements.txt

Restart, and the ~95 GiB INT8 ConvRot bundle into ComfyUI/models/MiniMax-H3-INT8-CONVROT/ (hf download Gluttony10/MiniMax-H3-INT8-CONVROT --local-dir ./models/MiniMax-H3-INT8-CONVROT, ModelScope in China). Legacy node; the modern RHMiniMaxH3VideoGen derives the target for you, but the knobs here are the same ones it's using. License as pack-wide: H3's Community License excludes the US, EU, UK and South Korea from local-weights use.