Architecture Knowledge HubArchitecture Knowledge Hub

Part 6 of 8: Your custom model trained. Your API can't see it yet.

By Pushparajan Ramar · September 23, 2026

Part 6 of 8: Your custom model trained. Your API can't see it yet.

Series: Firefly Services for developers. Checked against Adobe's documentation on September 23, 2026.

TL;DR

  • Train subject models for a specific product or character, and style models for a look and feel.
  • Both take 10 to 30 images in Firefly's training flow.
  • To call a model from the API, share it with your technical account first.
  • Then send an x-model-version header for custom models and pass customModelId in the body.

Subject or style?

  • Subject model: a specific character, person, product or object. Adobe's guidance: same make and model across images, subject in clear focus near the center, occupying at least 25% of the frame, varied views and lighting, and no large distracting objects. White or transparent backgrounds work, but a mix with more complex surroundings is best.
  • Style model: visual characteristics such as color palette, patterns, brush technique, illustration treatment, texture, lighting or overall aesthetic.

For subject models, Firefly asks for a Name and a Concept ID, a unique name you use in captions and prompts to refer to the subject.

Custom models are private by default and designed for commercial use. Access requires a Creative Cloud or Enterprise plan with Custom Models.

Training flow

Upload 10 to 30 images, let Firefly analyze them (it generates captions, model tags and recommendations for improving the set), correct the captions, resolve flagged issues, and train.

On data handling, Adobe states that its foundational Firefly models are not trained on your proprietary data.

The step people skip: sharing

Adobe's guide says a trained model must be shared with your technical account so that it's accessible to the List Custom Models API and the image generation API. Models shared at the organization level are also shared with individual projects.

flowchart TD
    A[Developer Console:<br/>open project, then API credential] --> B[Copy the Technical Account Email]
    B --> C[Firefly: open the custom model,<br/>... menu, Share]
    C --> D[Add the technical account email]
    D --> E[Model visible to the API]

Calling the model

Adobe's tutorial uses the async endpoint with a custom-model header value and the model's ID:

curl --request POST 'https://firefly-api.adobe.io/v3/images/generate-async' \
  --header 'Content-Type: application/json' \
  --header 'x-model-version: image3_custom' \
  --header "x-api-key: $CLIENT_ID" \
  --header "Authorization: Bearer $ACCESS_TOKEN" \
  --data '{
    "prompt": "An almond seed in a warm setting",
    "customModelId": "<your custom model id>",
    "numVariations": 2,
    "seeds": [66080, 82683],
    "size": {"width": 2048, "height": 2048},
    "contentClass": "photo"
  }'

The v3 spec's x-model-version values include both image3_custom and image4_custom. Adobe's tutorial demonstrates image3_custom, so confirm which your model supports.

Can't see your model? List them

curl -X GET https://firefly-api.adobe.io/v3/custom-models \
  -H "Authorization: Bearer $ACCESS_TOKEN" -H "x-api-key: $CLIENT_ID"

Models shared with your credential appear in the list. If yours is missing, go back to the share step.

Checklist

  • Subject and style trained as separate models
  • 10 to 30 images, same product and color for subjects, in clear focus
  • Technical Account Email copied from the right Developer Console project
  • Model shared with that email
  • Header and customModelId set on the request

Sources

Carousel

Download the PDF carousel