Inside look into AI inside Adobe Photoshop

Adobe Photoshop’s new generative fill and outpainting tools are delivering the kind of “paint-and-expand” results that can collapse hours of manual editing into minutes—while also exposing sharp limits around scale, server reliability, and content moderation. Hands-on testing shows the feature can seamlessly extend scenes, blend separate images into a single wide composition, and invent new elements (people, vehicles, animals, props) that match the surrounding context closely enough to feel integrated rather than pasted.

The most striking demonstrations come from outpainting—expanding an image beyond its original borders. In multiple examples, Photoshop extends a composition by filling in new space on the left, right, and even across large aspect-ratio changes. A time-lapse recreation of a scene blending two similar characters shows the tool can “stitch” the center and then keep expanding outward to form a coherent wide image, though it occasionally produces artifacts or lower-quality regions when the expansion gets too aggressive. Another test starts with a Midjourney-generated cyborg figure and expands it by selecting the bottom portion and using generative fill with a descriptive prompt; the system returns several candidate results, with the best option preserving the character’s overall look while filling the missing body area. When the expansion area becomes too large, the output can degrade into blankness, blobby shapes, or confusing white regions—suggesting the practical workflow is incremental, section-by-section growth rather than one massive outpaint.

Blending context across images is where the tool looks most impressive—and most fragile. When two similar cyborg images are placed side by side, Photoshop struggles with the boundary at first, producing a large confusing block. Iterative fills improve the connection, and the tool also appears to infer context cues such as gender, attempting to keep the second character’s traits consistent with the first. Still, the feature can fail when tonal differences are extreme or when the images are too close in ways that confuse the model’s understanding of where one scene ends and another begins.

The workflow also runs into operational and policy friction. During testing, Photoshop sometimes returns “system is currently busy” errors, which becomes especially frustrating because generative fill is integrated into a paid, professional product. More consequential are guideline enforcement moments: generated content can be removed after the fact for “potentially violates user guidelines,” even when the user is trying to do benign edits. Language strictness adds another layer of annoyance—misspellings or non-English prompts trigger corrections that reduce creative control.

Despite these issues, the practical takeaway is clear: generative fill is strongest for targeted edits, careful outpainting, and iterative refinement. Users can highlight specific problem areas (like a line of error) and regenerate until the seam disappears. The feature also supports “edits on edits,” enabling repeated passes to improve shadows, blending, and object placement. The testing ends with a playful but telling conclusion: the tool is fun, fast, and often surprisingly accurate, but it needs more reliability, more user control, and less overzealous moderation to fully match the promise of “best in painting” and “best in outpainting.”