Firefly Style Reference & Composition Reference

Adobe Firefly exposes two independent reference inputs that can be set from two different images, which is the core trick enabling free exploration of material/lighting without touching the actual shape:

Style reference: mark a previously generated (or uploaded) image via its pencil icon and "use as style reference" to define the material, lighting, and mood of subsequent generations. Generate a handful of options from a pure text prompt describing the desired material/environment/lighting (no shape involved yet), then pick the best one and set it as style reference before touching composition.
Composition reference: upload a shape/silhouette image in the composition section so the generated output follows that shape's structure. This is where the Photoshop-built shape (see Photoshop-to-Firefly 3D Render Pipeline) gets injected.
Strength slider: in the composition settings, controls how closely the final image must match the uploaded composition reference. Push it all the way up when the output must stay faithful to an exact shape (e.g. readable text or a specific silhouette); lower it if some AI-driven deviation from the shape is acceptable.

Firefly's shape/composition detection appears to treat touching same-colored regions as a single merged blob — this is why complex shapes need the fake-3D silhouette technique before upload, purely as machine-readable shape metadata rather than a visual choice.

Generated results can be regenerated indefinitely ("keep clicking generate after generate") and any result can be upscaled for higher resolution, at the cost of one additional credit.

Contradictions

None — new concept, no prior vault coverage of Firefly's reference-image mechanics.

Из тем: Beyond Photoshop: Other Tools & Prompting Techniques