Most image servers here generate from someone else's model. Being able to assemble a dataset and train your own from a conversation means the style becomes yours rather than borrowed — which matters when the output has to look like a brand. Note what the OAuth scopes are: identity only, so the authorisation screen is not where the capability boundary lives.
Ideogram's endpoint for their image product. The interesting third of it is not generation but training: you can build a dataset, upload assets to it and train a custom model from the same connection.
- generate_image and generate_images_bulk — from a prompt, with resolution and a model URI
- edit_image, remove_background, reframe_image, upscale_image — editing an existing image
- create_dataset, upload_dataset_assets, train_model — build a training set and train a custom model on it
- create_collection, upload_image, get_images_by_collection_id — organise the results
OAuth 2.0 with dynamic client registration and PKCE against their co-hosted authorisation server. The scopes requested are identity only — openid, email and profile.
Paste a URL, then authorize — add the endpoint to your client, then approve the OAuth consent screen