Firefly Custom Models API vs Sume: references, no training

Firefly's Custom Models API lets you train subject or style models and then pass a custom model asset ID to Generate Image. The Sume image docs describe no training step and no custom model ids. For brand consistency on Sume you pass reference images with each request: ChatGPT Image 2.5 takes up to 16.

Adobe facts are from its API overview; Sume facts from the Image API docs, read 2026-10-01.

What does Firefly Custom Models do?

Adobe says you can train subject models (characters, products, objects) and style models (color palettes, patterns, illustrative style). Each model has a unique asset ID for versioning and reuse, is hosted by Adobe, and can be shared with applications you grant access to.

What is the Sume equivalent?

Per request, not per trained model. input_references carries public HTTPS image URLs, and the model reads them along with your prompt. Models whose reference range is 0 to 0 are text-to-image only and reject references, so confirm the range in the catalog first.

Does Sume let me pass provider options?

No. The docs say allowed_passthrough_parameters is empty for every endpoint in v1, so provider.options must be omitted or empty. Parameters a model accepts are published as capability descriptors, and an unlisted one returns 400 unsupported_parameter.

What should I do for a consistent brand look?

Pick two to four clear reference images (the product, a style frame), write the constraints in the prompt, and reuse the same references for every call. This gives guidance, not a trained guarantee. For the nearest case on another vendor, see LoRA fine-tune ids and Sume.