AI is reshaping the photography workflow — from noise reduction to super-resolution, intelligent selection to generative fill. I have been using these features long enough to have opinions, and most of them are not about how impressive the results look in a demo. They are about how often the tools quietly get it wrong, and how much hand correction they actually leave me.
What AI already does well
The two genuinely solved problems are noise and masking. Lightroom's AI Denoise, added to Lightroom Classic in 2023, is the single best upgrade to my nocturnal work in a decade. I photograph Huaqiangbei's neon streets at night, and on a 5D Mark III at ISO 6400 the old tools smeared color into mush. AI Denoise cleans chroma noise while holding onto fine texture, and on the Sony Alpha A6000 files I shot years ago it revives images I had written off. The cost is real: it creates a new DNG, roughly doubling file size, it wants a decent GPU, and it takes seconds to tens of seconds per frame, so it is a culling-stage decision rather than something you run on an entire card.
AI masking is the other quiet revolution. Select Subject and Select Sky get me ninety percent of the way in one click, which changed how I dodge and burn. Where it fails is predictable and worth naming: hair against a busy background, translucent fabric, glass, and the soft edges of foliage still need manual brushing. I always refine the mask by hand before I trust it. The two seconds I save clicking are not permission to skip looking.
None of this is magic, and none of it stays still for long. Features get renamed, moved, and repriced between versions, so I try to depend on the outcome — a clean high-ISO file, a reliable mask — rather than on any single menu path staying where it is.
Where generative tools earn their place
Photoshop Generative Fill is a game-changer for removing unwanted elements (trash cans, tourists, power lines) — ten times faster than manual clone stamping. When I shot a market entrance I loved except for a stray signboard and two parked scooters, I selected them, typed nothing, and the frame was clean in seconds. Lightroom's Remove tool does the same job locally, and for small distractions it is honestly faster than a round trip to Photoshop. On Moraine Lake I used it to take a single hiker out of a rockpool reflection, and the replacement water texture was convincing at a glance.
Topaz Photo AI combines Denoise AI, Sharpen AI, and Gigapixel AI — stunning results for high-ISO night shots and enlarging old photos. But be careful: over-processing creates a "wax figure" effect. Back off the intensity for more natural results. My rule is to set denoise to roughly half of what the automatic suggestion proposes, then compare at 100% before committing. Stable Diffusion upscaling models (4x-UltraSharp) offer a free, controllable open-source alternative, though they require some technical setup — a capable GPU, a ComfyUI or Automatic1111 install, and patience with model files.
There is a practical cost most write-ups skip: Lightroom's Generative Remove and Photoshop's Generative Fill both spend cloud credits, so a heavy retouching session has a meter running. I keep a plain heal-and-clone fallback one keystroke away, and for anything small I use that instead of a generative pass, because it is free, instant, and predictable.
Adobe's own Super Resolution, hidden in the Enhance dialog, doubles the long edge of older files and is far easier than any of this. It works beautifully on well-exposed, sharp originals and does nothing useful for a soft frame, because it is a model of what detail usually looks like, not a record of what your lens actually captured.
Where it lies to you
AI does not restore missing information. It invents plausible information, and plausible is not the same as true.
The failures are specific. Generative Fill will invent foliage, brick, or ripples that look right and match nothing in the original — zoom in and you find surreal textures that never existed. It struggles with straight repeating lines, so railings and window frames come back warped, and it is hopeless with text and signage, which you should clone by hand instead. Reflections are a trap: extend a frame next to a lake and the new content rarely agrees with what the water is showing.
Denoise at strong settings erases exactly the texture you were trying to keep — feathers become frosting, pores disappear, and distant trees turn to paint. Sharpen AI can leave a bright halo along every high-contrast edge. And AI Lens Blur, which creates fake depth of field, still misreads complex edges, leaving a wobbling outline around hair and leaves that looks worse than the flat original. Upscaling never recovers true detail, only plausible detail; a soft image stays soft in substance no matter how crisp its pixels look.
My rules for using AI honestly
- Ask what the tool is allowed to do. Removing a tourist is cleanup. Generating a sky that was never there is fiction.
- Never tell a client an AI-edited feature was captured. AI should serve photographic authenticity, not replace it, and I always note AI modifications in image captions.
- Keep the original file. Every AI step stays reversible because I never overwrite the RAW, and I keep the pre-generative version as a separate layer or copy.
- Compare at 100% before export. Everything looks magical at fit-to-screen and questionable at full resolution.
- Start with the honest tools first. Exposure, white balance, and perspective correction are still the biggest wins, and no model fixes a badly shot frame better than a reshoot does.
Used carefully, AI gives me back hours of tedious work and lets me spend them on the decisions that actually matter — where to stand, when to press the shutter, and what to keep. Used carelessly, it produces images that look impressive and say nothing. The line between those two outcomes is not the software. It is the photographer's discipline about what the picture is allowed to claim.