AI video 4K quality prompt: what actually sets resolution

Words like “4K” or “ultra HD” in a prompt are not the documented way to get a 4K AI video. On Sume, resolution is its own request field: send resolution: "4K" to a model that lists 4K, and spend the prompt on the shot itself. One model lists 4K, gemini-omni-flash-1.1, and a finished clip can be upscaled instead.

The fields and limits come from Sume's Video generation, Video Router, and Timeline 1.0 docs and the Sume API reference, read on 2026-09-27. Anything described as current behavior is read from Sume's API code. 4K AI video generation API lists every model's resolutions.

Does putting 4K in the prompt make the video 4K?

Not when you pin a model. The docs define prompt as a text description of the video and resolution as the resolution of the output video, checked against the values the model advertises in supported_resolutions, plus the priced 2K and 4K upscales on minimax-h3. In the current code the resolution sent to the model is the field's value, and a value the model does not accept is refused with 400 unsupported_capability.

The bill follows the field too: Gemini Omni Flash 1.1 is billed per output second by resolution. With sume/auto, Sume picks the model family, and the documented defaults are 720p and 8 seconds, so pin the model when you need 4K.

What is the difference between native and upscaled 4K?

Native means the model generates the frames at that size. Upscaled means a smaller clip is generated first and a second pass enlarges it. On Sume, 4K comes three ways, and the docs compare none of them on detail, so judge the output yourself:

Should I regenerate a clip at 4K or upscale it?

Sume's tool guidance for agents points an existing clip to the upscaler and says: “Do not regenerate a clip to make it bigger.” A new generation is a new take, and with no seed on any v1 video model there is no documented way to get the same take back. An upscale works on the clip you already approved.

Video Upscale 1.0 (POST /v1/video-upscale-1.0/upscale) takes the clip as a public HTTPS video_url and enlarges it by a scale_ratio from 1.1 to 4, priced at $0.009 per video second on API pricing, plus a 5.5% agent fee by default. AI video upscaler API covers the request and its enhancement tiers.

What should a high-quality video prompt say?

Describe the shot. Sume's best-practice advice is specific, descriptive prompts with details about motion, camera angles, lighting, and scene composition. Choose the resolution to fit the use, since the docs say higher resolutions take longer to generate and cost more. A 4K request keeps the pixel size in its field:

{
  "model": "gemini-omni-flash-1.1",
  "prompt": "Slow dolly-in on a ceramic mug on a sunlit oak table, steam rising, shallow depth of field, warm morning light",
  "resolution": "4K",
  "aspect_ratio": "16:9",
  "duration": 8
}

What are the limits of 4K on Sume?

• Gemini Omni Flash 1.1 clips run 3–10 seconds, in 16:9 or 9:16 only.
• Video Upscale 1.0 reserves for at most 30 seconds of input: its duration_seconds field stops at 30.
• Timeline 1.0 output width and height are even integers from 256 to 2160, so a program joined there cannot keep a UHD 4K frame 3,840 pixels wide.
• In current code, a resolution a pinned model does not accept is refused before any job starts.