When Photoshop introduced generative fill in 2023, the internet declared the traditional retouching workflow dead. That turned out to be premature. What actually happened is more interesting: two distinct approaches now coexist inside the same application, and knowing when each earns its keep requires understanding what each one does to the pixels — not just what it promises on a feature slide.
This isn’t a verdict on which approach is superior. The classical workflow and the AI-assisted workflow are solving different versions of the same problem, and the trade-offs are specific enough to matter.
What the Classical Workflow Actually Involves
The classical Photoshop approach to a composite or retouching job follows a logic built around layer order, blending modes, and masking — every change made to specific pixel values is tracked through a layer stack that you can revisit, reorder, or discard. A skin retouching pass using frequency separation, for example, splits luminance (texture) information from color (tone) information into separate layers, letting you smooth tonal gradations on the low-frequency layer without disturbing pore detail on the high-frequency layer. Nothing is approximated; you are working directly with the underlying pixel data.
The same discipline extends to compositing. A selection made with the Pen tool defines a path that can be adjusted node by node. Refine Edge (or Select and Mask, in more recent versions) handles hair and semi-transparent fringe by sampling the edge pixels and generating a refined channel mask — a process that involves specific opacity values in the mask channel, not a probabilistic guess. You can audit that mask at any stage by Alt-clicking the layer thumbnail.
This is labor-intensive. A complex hair mask on a textured background might take twenty to forty minutes from rough selection to a channel refined enough for print. That time cost is the central argument the AI tools have been making against the classical approach.
What the AI Tools Are Actually Doing Differently
Adobe’s AI features — Generative Fill, Generative Expand, Remove Tool, and the AI selection refinements — operate through a fundamentally different mechanism. Rather than letting you define pixel-level relationships manually, they sample context from the surrounding image and synthesize new content using a diffusion model trained on a large image dataset. The output pixels were not in the original file; they are statistically plausible given the neighborhood.
This distinction has real consequences. When you use Generative Fill to extend a background, the generated content is inserted as a new layer containing synthesized pixels. Those pixels are not derived from anything in your original scene; they are inferences. Whether the result reads as photographic depends on the complexity of the scene — a plain studio background extends credibly, a detailed architectural scene with consistent perspective is harder to fake without visible tell-tale softness or repeated texture motifs.
The Remove Tool for object removal similarly synthesizes replacement content rather than cloning from a user-specified source. For patches surrounded by consistent texture — grass, pavement, neutral sky — the synthesis is often cleaner and faster than a careful content-aware fill or clone stamp pass. For patches near strong geometric edges, the tool can introduce subtle perspective errors or tonal seams that the classical healing brush, guided by a user-specified sample point, avoids.
Select Subject and the AI-driven hair refinement have improved meaningfully in recent versions. They are operating on learned edge patterns rather than pixel-contrast heuristics alone, which helps in cases where the subject and background share similar luminance values. The resulting mask is still a grayscale channel, but the values in it were computed by a model rather than by your explicit brush strokes.
Comparing the Two Approaches on Specific Tasks
The practical differences sharpen when you put both approaches against the same image problem:
| Task | Classical Approach | AI-Assisted Approach |
|---|---|---|
| Background removal (clean studio shot) | Pen path or Select Subject + manual refine | Select Subject + AI mask refinement; fast and often sufficient |
| Background removal (complex hair, outdoor) | Channel-based mask, often 20–40 min | AI refinement faster but may require manual cleanup at fine edges |
| Object removal (surrounded by texture) | Clone stamp or content-aware fill | Remove Tool often cleaner, requires less sample-point management |
| Object removal (near geometric edges) | Clone stamp with manual perspective attention | Generative synthesis can introduce errors; classical may be safer |
| Tonal retouching | Curves, levels, dodge/burn on dedicated layers | AI-powered tools exist but tonal work remains primarily classical |
| Extending a frame edge | Content-aware scale or careful cloning | Generative Expand synthesizes plausible extensions; quality varies by scene complexity |
One pattern emerges clearly: the AI tools front-load speed and reduce the skill ceiling for routine tasks. But they also reduce transparency. A mask generated by Select Subject does not tell you why it drew the boundary where it did. A Pen path is completely auditable. Generative Fill produces pixels you cannot reverse-engineer — you can undo the operation, but the synthetic output carries no record of its own construction.
For documentary or archival work, that opacity carries real implications. The question of what counts as a manipulation, and what a file actually contains, becomes harder to answer.
Where the Hybrid Approach Makes Sense
Most experienced editors working in Photoshop today use both paradigms within a single session — and that turns out to be the more practical framing than treating them as a binary choice.
A portrait retouch might begin with Select Subject’s AI mask to establish the rough separation, then move immediately to a channel-based refinement pass on the flyaway hair, then use classical frequency separation for the skin, then lean on the Remove Tool for a distracting background element surrounded by consistent texture. Each tool is used where its mechanism suits the problem.
Photoshop’s Light Layer Looks Like Lightroom. It Doesn’t Act Like It. addresses a related situation where a new Photoshop feature’s surface similarity to a familiar Lightroom concept can mislead an editor about what it’s actually doing to pixel data — which is the same trap here. Knowing that AI selection refinement is generating mask values from a model rather than sampling actual edge contrast in your file changes how carefully you verify the result before committing.
Newer AI features like editable AI masks in Lightroom operate on a similar model, and if you’re already familiar with how those work — or how they can fail at subject boundaries — you’ll recognize the pattern when auditing an AI-generated mask inside Photoshop. Our article on Lightroom’s new editable AI masks covers the mechanics of that specific implementation.
What You Should Verify Before You Commit
Whichever approach shapes a given session, the verification steps that matter most are often the same:
- Zoom to 100% on critical edges — AI-generated masks frequently look clean at fit-to-screen zoom and reveal fringing or misclassified edge pixels only at 1:1.
- Alt-click the layer mask before flatting or exporting to inspect the actual mask channel values rather than the composited preview.
- Check generative content against perspective and lighting direction — synthesized patches that ignore the scene’s light angle are easier to catch by eye than the tool’s confidence level suggests.
- Preserve your layer stack until final output — converting to a flattened file before client approval removes your ability to revisit AI-generated selections if a problem surfaces in review.
- On archival or editorial work, document the tools used — if your output will be subject to disclosure standards or contest rules, AI-generated pixel synthesis and manual retouching carry different implications; knowing which layers contain synthesized content matters later.
The question to carry into any editing session is not “should I use the AI tool?” but “do I understand what this tool did to this specific region of this file, and can I verify the result?” That question has always been the right one in classical Photoshop work. It turns out to apply to the AI tools just as firmly — the answer is just harder to find.