Headshots from photos

Produce the requested professional portraits of the supplied person with Astria. Preserve identity and natural features while changing the requested wardrobe, background, lighting, or framing. Do not replace them with an invented face.

Use astria-api for schemas, reference syntax, HTTPS inputs, workspace scope, model discovery, idempotency, and checking completed results.

Establish the identity reference

Use an explicit saved reference; resolve a named identity with list_references and inspect it using get_reference. Otherwise create one identity reference with create_reference from the supplied usable HTTPS photo URLs, using a subject class appropriate to the supplied brief. Use the returned ID and class name on every portrait. Do not create a new reference for each background.

For an attachment or local file, use a real host upload capability or the optional authenticated CLI. If neither can supply accessible URLs, ask for usable media instead of inventing a URL. If no identity input exists, ask for photos or a saved reference rather than generating a stranger.

For a team, keep each person's photos and reference ID separate; a group of faces is not one identity. Preserve the requested headcount and shot count.

Generate the portraits

Use a requested saved template through get_template / run_template, resolving its actual reference slots before submission. Otherwise call list_models and use the selected model or catalog default with generate_images.

Include the returned identity token and class in the text, then describe the requested professional style. For an unspecified style, choose simple studio lighting, a neutral background, natural expression, and professional clothing. Make one image unless the user asks for a set. Use separate prompts for distinct backgrounds or outfits, reusing the identity reference throughout.

Do not borrow avatar's invented age, heritage, facial traits, or casting prompt. Do not infer protected attributes from the identity photos.

Deliver

Retain returned prompt IDs and check get_prompt for each requested output. Present completed images and report pending or failed portraits accurately. When inspection is available, check resemblance, natural facial proportions, and framing. Report an identity mismatch without claiming an exact likeness or automatically spending credits on additional attempts.