4K AI image generation API: resolution, quality, and image count

To generate a 4K image with Sume's API, send POST /v1/images with resolution: "4K" to google/nano-banana-2 or google/nano-banana-pro, the only catalog models that list a 4K tier. The ChatGPT Image models reach a 3840-pixel edge through custom sizes instead, and every model publishes its own quality levels and per-call image count, n.

Tiers, levels, and ceilings below are read from the descriptors GET /v1/images/models serves and the Image API docs, on 2026-09-26.

Which image models support 4K?

resolution is a normalized tier (512, 1K, 2K, 4K), and size is shorthand for the same tiers. Only google/nano-banana-2, google/nano-banana-pro, and google/imagen-4-ultra list tiers at all, and Imagen 4 Ultra stops at 2K. On any other model, a tier on either field returns 400 unsupported_parameter. Every output control, model by model:

How do I request a 4K image?

On either 4K model, send resolution: "4K" or size: "4K" with a ratio from its list. Leave both out and the image renders at 1K. That family's ratios and edit flow have their own guide.

The ChatGPT Image models list no tier. Their largest frame is a custom image_size such as 3840×2160, which sits exactly on the 3840-pixel edge and 8,294,400-pixel ceilings; the GPT Image 2.5 API has the full size rules. Explicit pixels on size return 400 unsupported_parameter.

curl -X POST "https://api.sume.com/v1/images" \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "google/nano-banana-2",
    "prompt": "a wide mountain landscape at dawn",
    "aspect_ratio": "16:9",
    "resolution": "4K"
  }'

Which quality levels can I set?

quality is catalog-gated. ChatGPT Image 2.5 and 2.5 Sunburst take auto, low, medium, high, xhigh, and max; ChatGPT Image 2 and Ideogram V3 take low, medium, and high. The three ChatGPT ids run at high when you omit the field. Every other model lists no quality, so sending one returns 400 unsupported_parameter.

On ChatGPT Image 2.5, the level also moves the estimated price, as the GPT Image 2.5 API explains.

How many images can one call return?

The request schema allows n up to 10, but each model's n range is lower: 4 on most rows and 1 on Grok Imagine (x-ai/grok-image). A value above a model's ceiling is refused with a 400 error. Seedream 4.5 lists 4, yet a Seedream 4.5 request with reference images returns one image.

output_format is png, jpeg, or webp everywhere except Recraft V4, which lists WebP only, and a Seedream 4.5 request with references does not apply it. svg is in the schema, but no model lists it, and output_compression returns 400 unsupported_parameter.

Why did my 4K request return 202?

POST /v1/images blocks for up to 30 seconds. 4K, high quality, and large n are the configurations most likely to run past that budget and come back as 202 with a job envelope. Branch on the status code: poll GET /v1/jobs/{id}/status, then fetch GET /v1/jobs/{id}/result, or send mode: "async" from the start.

How do 4K and extra images change the price?

Endpoint pricing is per output image, so every image in n adds a charge. On the two 4K models the tier can change the per-image rate too, and the catalog line shows the 1K rate; their guide lists which tiers change it on each.

Catalog lines already include Sume's margin, plus a 5.5% agent fee by default; rates are in GET /v1/images/models and on API pricing. A failed or cancelled generation is not billed.