Can Imagen 4 use my product photo? No: pick an edit-capable row

No: Imagen 4 Fast and Imagen 4 Ultra on Sume are text-to-image only, and the catalog lists input_references as 0 to 0 for both, so a request that attaches your product photo is rejected with 400 unsupported_parameter. For a brief that starts from a photo, pick one of the edit-capable rows: Nano Banana 2.1, Seedream 5.0 Lite, FLUX.2 Pro, Grok Imagine, Qwen Image, ChatGPT Image 2.5 or Ideogram 4.5.

Google's own Imagen page, read on 2026-10-10, says Imagen is shut down in the Gemini API and points developers to Nano Banana models. That is a statement about Google's API. This post is about what the Sume catalog lists and accepts, taken from the Image API docs and the catalog code.

What the Imagen rows do accept

Both rows take a prompt, n from 1 to 4, one of five aspect ratios (1:1, 16:9, 9:16, 4:3, 3:4) and an output format of png, jpeg or webp. Imagen 4 Ultra also lists resolution of 1K or 2K. Prices billed on Sume are $0.025 for Fast and $0.075 for Ultra.

The Sume docs state the rule directly: if the input_references descriptor of a model is {"min": 0, "max": 0}, the model is text-to-image only and rejects references. They also say Sume does not silently drop a parameter the row does not list.

Rows that take a product photo

Every other edit-capable row lists up to 10 references, except ChatGPT Image 2.5 at 16 and Ideogram 4.5 at 5. Prices are the billed catalog price per output image at the default tier.

A workable ad-creative path

If you used Imagen for cheap drafts and now need the product in frame, keep the draft step on a text-only row and move the finish to an edit row. For example, 40 drafts on Imagen 4 Fast cost 40 x $0.025 = $1.00. Ten product-photo edits on Nano Banana 2.1 at 1K cost 10 x $0.10 = $1.00. That is $2.00 for 40 concepts and 10 finished edits.

Keep the photo URL public HTTPS and pin the model id. The Sume sume/auto value hides which family ran, so it is a poor choice when a brief depends on a reference being honored.

• Read input_references in the endpoints route before you wire a new row into a pipeline.
• Use aspect_ratio: "auto" on edits only where the row lists it (Nano Banana, ChatGPT Image, Seedream 4.0).
• Test one edit before you queue a batch.

Reading the 400 when it happens

When a request attaches input_references to an Imagen row, the response is a 400 with the unsupported_parameter code, and the request is not billed because nothing was generated. Read the error, drop the references or change the row, and resend. Do not retry the identical request, because the answer will not change.

A cheap guard is to read supported_parameters.input_references from GET /v1/images/models at startup and keep a set of text-only ids in your config. The same call tells you each row's max for references, so one lookup serves both the edit and the text-only branches of a pipeline.

One more point on naming: Sume still accepts bare Image Router ids such as gpt-image-2 as aliases for their org/slug equivalents, but new code should send the google/imagen-4-fast form that the catalog lists. Whichever form you use, the response echoes the id that you requested.

A short checklist before you migrate

Start by listing every call that still sends an Imagen 4 id. For each one, note whether it sends reference images: if it does not, the call is a plain text-to-image job and any text-capable row can replace it. Next, note the ratio the call asks for, because Imagen 4 rows list only five ratios and every row in the edit table lists at least those five.

Then run the same three prompts on the candidate row and on the old row. Compare the output side by side, and only then change the default. Keep the old id in config for one release so you can roll back, and log the row id with every result so a regression can be traced to a model change rather than a prompt change.

Finally, re-check the price. Imagen 4 Fast is billed at $0.025 and Imagen 4 Ultra at $0.075 on Sume. A replacement at $0.10 costs four times the Fast row, so a high-volume draft job may want a cheaper row such as FLUX.2 Pro or Qwen Image, and a final-quality job may want a more capable one.