MiniMax H3 Fun ControlNet: 17n+5 frames vs Sume whole seconds

For the local MiniMax-H3-Fun-Controlnet-Union-2.0 checkpoint, the model card says the frame count snaps down to the largest 17 * n + 5 the video VAE can decode, at a fixed 24 fps and a 15-second cap. On Sume there is no frame field: you send duration in whole seconds and the model catalog lists the allowed values.

Control-model facts are from the model card, read 2026-10-01. Sume facts are from Video generation, read 2026-10-01.

What frame counts does the ControlNet card allow?

The card says the generation follows the control video. The frame count snaps down to the largest 17 * n + 5 the video VAE can decode, and the duration is capped at 15 seconds. The canvas keeps the control video's aspect ratio within a pixel budget, with both sides multiples of 32, and the frame rate is fixed at 24 fps.

The table below is arithmetic on those rules, assuming the control video has the frame count shown at 24 fps.

How does Sume describe duration?

The catalog field supported_durations is described as supported video lengths in whole seconds, so the answer to "how many frames" is not part of a Sume request. The same page lists minimax-h3 at 5-15 seconds at native 480p/768p, and names a few models that run longer (seedance-2.5 and wan-3.0 up to 30 seconds); every other catalog model tops out at 15 seconds. Read the numbers per model from GET /v1/videos/models before you submit.

Should I convert frames to seconds myself?

Yes, if you drive a local control model and a hosted one from the same plan. Work in seconds for the hosted request and round down for the local one: a 10-second plan becomes duration: 10 on Sume and 226 frames locally, which plays for about 9.42 seconds. If the two outputs must match exactly, trim the longer one afterwards. For other frame rules see LTX frame counts.

Does Sume run the ControlNet branch?

Nothing in the cited Sume docs lists a ControlNet or control-video input for minimax-h3. The docs say minimax-h3-max supports text-to-video, first/last-frame image-to-video and reference-to-video. Use the catalog to see which supported_input_references a model accepts; the local checkpoint's eight control conditions are a separate workflow, covered by the card and by MiniMax H3 in ComfyUI.