Five Sume image models are text-only. What to use with a photo

Five rows in the Sume image catalog are text-to-image only: higgsfield/soul, google/imagen-4-fast, google/imagen-4-ultra, recraft/recraft-v4 and qwen/qwen-image-max. Their input_references descriptor is {min: 0, max: 0}, so a request with a reference image is rejected, not ignored. If you have a photo to edit, use an edit-capable row; the cheapest of them is $0.025 per image.

This comes up most after a launch week. Recraft's blog lists V4.1 and a faster V4.1 Flash, and a reader sees a new name and assumes a reference field. Sume's Recraft row is V4, text only, and it returns webp.

The five text-only rows

The swap table

Pick the edit-capable cousin that is closest in style or price.

• Qwen Image Max with a photo: use qwen/qwen-image at $0.025, up to 10 references.
• Imagen 4 with a photo: use google/nano-banana-2.1 at about $0.10 or Seedream 5.0 Lite at about $0.04375.
• Recraft V4 with a photo: use FLUX.2 pro at $0.0375, which lists the same 13 ratios and takes 10 references.
• Soul with a photo: use Seedream 4.5 or Grok Image; Soul has no edit path.

Check before you send

Read the descriptor instead of memorizing this list: supported_parameters.input_references.max is 0 on text-only rows. A short pre-flight in code is cheaper than a failed request, and it keeps working when the catalog changes.

If you must keep the look of a text-only row, generate with it first and use the output as the reference for an edit-capable row. Two calls cost the sum of both prices, for example $0.05 plus $0.0375 for a Recraft V4 image refined on FLUX.2 pro, which is $0.0875.

Reading the catalog in code

A short pre-flight avoids paying for a failure round trip. Fetch GET /v1/images/models/{id}/endpoints, read supported_parameters.input_references, and branch on max. Cache the result for the length of one job, not for weeks, since the catalog moves when providers change.

If you build tooling for a team, show the swap table above next to the model picker so nobody has to learn the rule by hitting a 400. The cost of a wrong guess is time, not money, because a rejected request is not billed.

What Sume does not do

Sume does not silently drop a reference to make a text-only request pass, and it does not route a text-only request to another model. Pick the row, and read the error if the catalog says no. Imagen availability can change, so confirm the row is in GET /v1/images/models today.