Which Sume image models are text-to-image only, with no photo edits

Five image models in Sume's catalog are text-to-image only and cannot edit a photo you send: Higgsfield Soul, Imagen 4 Fast, Imagen 4 Ultra, Recraft V4 and Qwen Image Max. The other fifteen models in the catalog, including every ChatGPT Image, Nano Banana, Seedream, FLUX.2, Grok Imagine, Qwen Image and Ideogram row, accept reference images. If you need to change an existing picture, use one of the fifteen.

The catalog descriptor shows it as an input_references range with a maximum of 0; edit-capable models list a maximum of 10, or 16 on ChatGPT Image 2.5. For Recraft V4 the constraint text is explicit: text-to-image only; image URLs are not accepted. Send a reference to one of these five and the API rejects the call rather than ignoring the image.

The five, with price and ratios

Billed prices below are list times 1.25. All five take up to 4 images per call. Soul's price is for 720p, with 1080p as the other tier.

What to use when you need an edit

The cheapest edit-capable row is Grok Imagine and Qwen Image at $0.025 billed, followed by Seedream 4.0 at $0.0325 and FLUX.2 Pro at $0.0375. For a high-fidelity edit, ChatGPT Image 2.5 takes references and also offers quality up to max. Pick on price from that list and test on your own images.

A common mistake is to pick a text-only model for a workflow that starts as text and later needs an edit step. A product-photo pipeline that renders a background and then wants a retouch will hit 400 unsupported_parameter on the second call. If a pipeline may ever need an edit, choose an edit-capable id up front so the model does not change halfway through.

For a text-only model, the nearest edit-capable neighbour by price is a sensible first test. Recraft V4 at $0.05 pairs with Seedream 4.5 at the same $0.05; Imagen 4 Ultra at $0.075 pairs with Ideogram 4.5 at its medium tier. Both neighbours accept references.

Why these five are text-only

The capability record for each of these lists no reference range, and the API enforces the catalog. A mixed workflow therefore needs two model ids. Generate with the text-only model, and if the second step is an edit, send the generated image's URL as a reference to an edit-capable model. The response carries a URL for each image rather than base64, so the first result can be passed straight on as a reference, as long as the URL is reachable by the second call. A reference-based second step costs the edit model's normal per-image price.

Price per edit-capable alternative

To keep the swap concrete, the table-free version is this. Soul at $0.005 has no edit sibling at that price, so a Soul user who needs edits moves up to Grok Imagine or Qwen Image at $0.025, five times the price. Imagen 4 Fast users pay the same $0.025 on Grok Imagine or Qwen Image, so for them the swap is free. Recraft V4 users pay $0.0375 for FLUX.2 Pro, which is less. Qwen Image Max users, at $0.09375, pay less on most edit-capable rows; the exceptions include Nano Banana 2.1 and Pro at 1K and the highest ChatGPT Image tiers.

How to check any model in code

The list endpoint returns each model's capabilities, so you do not have to trust a blog post for this. The command below prints the ids whose input_references maximum is 0, which are the ones that cannot edit. If Sume adds an edit-capable variant of one of the five, this will show it before this page is updated.

curl -s https://api.sume.com/v1/images/models \
  -H "Authorization: Bearer $SUME_API_KEY" \
  | jq -r '.data[] | select(.supported_parameters.input_references.max == 0) | .id'