Which AI video models take 1080p on Sume, and which do not

Seedance 2.5, the Seedance 2.0 rows (seedance-2, seedance-2-fast, seedance-2-mini), kling-3, wan-3.0 and gemini-omni-flash-1.1 all accept resolution: "1080p" as a plain option. minimax-h3-max also accepts 1080p, but as a latent refinement of a native 768p render under the same model id. minimax-h3, grok-imagine-video-1.5 and higgsfield-genjutsu do not offer 1080p at all.

The matrix comes from Sume's Video Router docs, Video generation docs and the OpenAPI resolution description, read on 2026-10-03. It tells you what a request is allowed to ask for. It does not tell you how much detail a model resolves, which only your own clip can show.

Which video models take 1080p?

Read the third column carefully. Sume's docs call out one case explicitly: for MiniMax H3 Max, 1080p is a latent refinement from native 768p and uses the same model. The docs say nothing equivalent about the other rows, so this post does not claim that they render natively at 1080p either; it only reports what the request accepts.

Resolution is also not independent of length. Omni accepts 3 to 10 seconds at every resolution, while Seedance 2.5 accepts 4 to 30 seconds at 480p, 720p and 1080p. A long 1080p clip is a different price from a short 720p one, and the per-second rate is by resolution on the models Sume prices that way.

How do I see what I actually got?

Ask for 1080p, then probe the file. Sume's video inspect takes one media.sume.com clip, such as the artifact URL on your finished job, and returns probe facts plus stills with their width and height. Compare those against what you requested. It also lets you pull a still at an exact timestamp to check detail on the face or the label, which is the thing 1080p was supposed to buy you.

A frame of evidence beats a spec sheet. A file can report a 1920 by 1080 frame size and still look soft, so judge detail from stills, not from the container.

• Run video inspect on a draft and a final and compare stills at the same timestamp.
• Do not assume 1080p on Kling 3 is better than 720p for your subject; test both.
• Use 480p drafts where a model offers them, and reserve 1080p for the take you will ship.
• Pin the model, because sume/auto creates 720p clips by default and does not name the model it used.

When is 1080p not enough or too much?

If a platform wants 4K, only gemini-omni-flash-1.1 accepts native 4K in the Video Router; 4K AI video generation API lists the resolutions model by model. For MiniMax's own 2K and 4K, those are priced upscale SKUs on minimax-h3, not on H3 Max.

If you only need a draft, 480p on the Seedance or Wan rows is the cheaper way to iterate, and 720p is the default when you omit resolution on most rows. The stored post on H3 Max 1080p latent refinement goes deeper on that one case.