What Generative Fill does
Generative Fill uses Adobe's Firefly model to create new pixels based on a text prompt and the surrounding image. It runs inside Photoshop 2024 and newer, lives on the contextual taskbar and writes its result onto a new Generative Layer — so the original pixels stay intact.
Step 1 — Extend a background
Press C for Crop, drag the canvas outward, hit Enter. Make a rectangular selection across the new empty area plus a small strip of existing image. Click Generative Fill on the taskbar, leave the prompt blank and click Generate. Photoshop fills the area seamlessly in three variations.
Step 2 — Remove unwanted objects
Lasso loosely around a person, sign or wire you want gone. Click Generative Fill with an empty prompt — leaving the prompt empty tells Firefly to extend the surroundings instead of inventing new content. Result is often cleaner than Content-Aware Fill.
Step 3 — Add new objects
Select an area where you want something added, type a short prompt — "vintage red bicycle leaning on a wall" — and click Generate. Stay specific about lighting and material; Firefly understands terms like "golden hour", "matte" or "film grain".
Step 4 — Refine variations
Each generation produces three options in the Properties panel. Click through the thumbnails or press Generate again for fresh variations. Mask any unwanted areas on the Generative Layer instead of regenerating — masking is free and instant.
Step 5 — Match grain and color
AI-generated pixels can look too clean. Add a Camera Raw filter on a stamp-visible layer and dial in grain 8–12 and noise 3–5 to match the original capture. Apply a subtle Curves layer to unify tone across original and generated regions.
Best practices
Generate at the smallest selection that includes 10–20 pixels of context. Keep prompts under 12 words. For commercial work, save the Edit History (File → Generate Image History) — Adobe records Content Credentials for every generation, which clients and stock agencies increasingly require.
Limits
Firefly will refuse known IP, real people's faces and unsafe content. Output is capped at 1024×1024 per generation in current versions, so for large fills work in tiles or use Generative Expand.