Photoshop Firefly Fill and Expand vs Firefly Image 3: What Actually Improved?

Adobe has added a new generative model to Photoshop called Firefly Fill and Expand, and it directly affects how you create, replace, and extend images. If you rely on generative fill for background swaps or composite work, these changes are worth a look.

Coming to you from Anthony Morganti, this practical video compares the older Firefly Image 3 model with the new Firefly Fill and Expand model inside Photoshop. Morganti starts with something many people struggle with: generating a believable person from scratch. Using the same prompt in both models, he shows how the older version still produces warped hands, distorted faces, and strange background details. The new model does better, especially with hands and facial structure, and the files come out at twice the resolution. Even so, results remain inconsistent, and you still see odd artifacts and unusable variations.

That matters when you attempt to build a scene that includes people, because you cannot depend on clean output every time. One run might look acceptable, the next might fail in obvious ways. The newer model produces sharper files at 2K instead of the earlier 1K output, and that extra resolution shows up in finer detail. But clarity alone does not fix structural issues. If you plan to use generative tools for client-facing composites, you need to know how far you can push them before flaws become visible.

Where the new model clearly improves is background replacement. Morganti demonstrates this with a zoo portrait of a gorilla, selecting the subject and asking Photoshop to replace the background with a forest. With Firefly Image 3, the background swap often alters the edges of the subject, subtly reshaping hands, fur, and facial contours. The newer Firefly Fill and Expand model protects the original subject more faithfully. The edges remain more intact, and while minor changes still happen, they are reduced. You also get access to a detail enhancement option that can refine the generated area, which helps when textures look slightly soft.

The video also covers reference-based swaps, where you replace a selected area using another image instead of a typed prompt. This feature is unavailable in the older model but active in Firefly Fill and Expand. Morganti demonstrates swapping a woman’s head using a second portrait as the reference. The system offers options such as referencing only the object or the whole image, and different intent settings for how the replacement behaves. The results are usable at a glance, though hair edges and blending still require scrutiny. It works, but you would not call it seamless without cleanup.

Finally, Morganti tests generative expand by cropping outward on an image and allowing Photoshop to fill new space automatically. The difference shows up most in texture continuity. Sand, for example, blends more naturally and with better detail in the newer model. When you zoom in, the transitions appear cleaner compared to the softer, blurrier output from Firefly Image 3. It is subtle until you inspect closely, then the gap becomes obvious. Check out the video above for the full rundown from Morganti.