Photoshop just updated Generative Fill with the new Firefly Fill and Expand model, and yes, it changes what you can realistically create. If you rely on AI inside Photoshop, this affects how large you can generate, how real people look, and whether hands and cars still fall apart.
Coming to you from Colin Smith with photoshopCAFE, this detailed video walks through what actually changed inside Generative Fill and how the new Firefly Fill and Expand model compares to Firefly Image 3. The biggest shift is resolution. The old model topped out at 1,024 by 1,024 pixels, or 1K. The new model doubles each dimension to 2,048 by 2,048 pixels. That jump means you can generate over much larger areas before you start seeing softness and stretched detail. On a 1,920 by 1,080 document, the new model fills the frame at full resolution, which reduces the muddy look that used to appear when pushing bigger selections.
That resolution boost changes how you work. With Firefly Image 3, you had to build in smaller sections to protect detail. Large selections forced the model to stretch pixels, which made edges soft and textures break down. With Firefly Fill and Expand, extending a frame using the Crop tool and Generate Expand produces cleaner transitions and more believable grain. At closer zoom, you still see structure fall apart, so large print work remains questionable, but for web, social, and most multimedia projects, it is far more usable. The video shows side-by-side comparisons that make the jump obvious without overselling it.
Generative credits stay the same for standard use: one credit per generation, and that includes three variations. There is no refund if the result misses the mark. Other models inside Photoshop cost significantly more credits per generation, but Firefly Fill and Expand remains at the base rate. If you experiment heavily, the cost structure may influence which model you test first.
Where this update gets interesting is in how it handles people. Using the same prompt on both models, the difference is clear. Firefly Image 3 tends to create soft, slightly cartoon-like faces with weak detail. Skin can look plastic, and proportions drift. The new Firefly Fill and Expand model produces faces that are sharper and more believable at normal viewing size. Full-body generations in studio lighting show better anatomy overall, though hands still give it trouble. Thumbs can look oversized, and fingers sometimes appear fused or overly smooth. The improvement is real, but not flawless.
The video also tests more complex prompts, placing a person in a jungle with cinematic lighting. Firefly Image 3 struggles with realism in both clothing and anatomy. The newer model generates stronger proportions and more convincing lighting, even if fabrics sometimes look slightly illustrated. You see progress, yet you also see where edges remain rough.
Hands remain a stress test. Firefly Fill and Expand produces more accurate five-finger anatomy than before, but detail can swing too far in either direction. The older model over-textured skin. The newer one sometimes looks waxy, with fingertips that resemble molded clay. You gain structure but still lose subtle realism.
Cars are another weak point. The old model produces mangled shapes, strange wheel counts, and unrecognizable designs. The new model gets the number of wheels right and suggests familiar silhouettes, though sharpness and brand-level detail are still lacking. It improves perspective and coherence, but it does not match higher-end generative tools outside Photoshop.
One feature mentioned briefly stands out: the ability to use reference images inside Generative Fill. That opens up more controlled results, especially when style or consistency matters, and the video promises a separate walkthrough on how to use it properly. Check out the video above for the full rundown from Smith.
No comments yet