The five things AI cannot fix in a source photo
Generative editing is additive and transformative, not archaeological. When information is destroyed at capture, the model fills the gap with plausible invention — which is exactly where fake-looking results and product inaccuracies come from. Five defects are effectively permanent:
| Defect at capture | What AI does with it | Fixable? |
|---|---|---|
| Blown-out highlight (clipped white) | Invents texture that was never there | No |
| Out-of-focus product edge or label | Hallucinates letterforms and stitching | No |
| Low resolution (under ~1000px) | Upscales, softening fine detail and pattern | Rarely |
| Cropped product (part out of frame) | Guesses the missing geometry | No |
| Heavy motion blur | Smears into a painterly texture | No |
| Cluttered or busy background | Masks and replaces it cleanly | Yes |
| Flat, dull lighting | Relights convincingly | Yes |
| Wrong background color | Swaps it entirely | Yes |
| Dust, lint, minor scuffs | Retouches away | Yes |
| Mild wrinkles in fabric | Smooths or re-drapes | Yes |
Read the table as a division of labor. Everything in the green column is work you should stop doing manually — background, staging, and cleanup are precisely what AI handles well. Everything in the red column is work only your camera can do, and it takes minutes.
White-background shooters routinely overexpose to "get a clean white." This clips the product's own highlights too. Shoot the product correctly exposed against a grey or mid-tone surface and let AI produce the pure white background — you keep every highlight detail.
Resolution and framing: the numbers that matter
Marketplace requirements set the floor, but the working floor for AI input should be higher than your output target, because any reframing, cropping, or perspective change consumes pixels.
Note the 85% figure is a capture guideline, not an output one. Amazon wants the product to occupy roughly 85% of the finished frame, but you should shoot looser than that — leave breathing room on all four sides. Generative reframing can crop in without penalty; it cannot extend a product that runs off the edge without inventing geometry.
On file format: shoot the highest-quality option your camera offers. Phone HEIC or ProRAW both work. Avoid feeding heavily compressed JPEGs pulled from a supplier catalog or a social post — the compression artifacts get amplified, not smoothed, and they read as a grainy halo around product edges in the finished image.
Lock focus and exposure before you shoot. On iPhone, tap-and-hold on the product to lock AE/AF, then drag down slightly to protect highlights. One tap eliminates the two most common unfixable defects at once.
Lighting: soft, even, and boring is the goal
Counterintuitively, dramatic lighting is the wrong thing to capture. Hard directional light bakes shadows and specular hotspots into the pixels, and every downstream restaging inherits them. Flat, even light is a neutral canvas — AI can add drama later, but it cannot remove a shadow that is physically fused to the product surface.
Shooting for a final image
- Dramatic side light and deep shadow
- Colored gels for mood
- Reflections staged in-camera
- Background lit and styled on set
- Every scene needs a re-shoot
Shooting for AI input
- Soft, even light from a large source
- Neutral white balance, no color cast
- Specular hotspots minimized
- Plain background, product isolated
- One capture feeds many scenes
A north-facing window with a white sheet diffusing it, plus a white foam board as a fill reflector opposite, produces source photos that outperform most speedlight setups for this purpose. The aim is a full tonal range with nothing clipped at either end: no pure-white blowouts, no crushed blacks swallowing the product's silhouette.
White balance deserves particular attention because it is the one lighting error that is technically correctable but practically costly. A mixed-light shot — daylight from a window plus warm tungsten overhead — produces a color cast that varies across the product surface. Correcting it globally leaves one region wrong. Turn off the room lights and use a single light family.
How many angles to capture, and which ones
Every additional source angle multiplies what AI can generate downstream, and the marginal cost of one more frame is a few seconds. The mistake is shooting one hero angle, discovering later that the lifestyle scene needs a three-quarter view, and having to unbox the sample again.
A practical baseline for most physical goods:
- Straight-on front. The reference frame for hero images and marketplace main images.
- Three-quarter. Rotated roughly 45 degrees — the most useful angle for lifestyle scenes because it reads as dimensional.
- Top-down. Essential for flat lays, apparel, and anything where footprint matters.
- Detail macro. One tight frame on the material, texture, stitching, or finish. This is what buyers zoom into.
- Back or underside. Labels, care tags, ingredient panels, serial plates.
For apparel specifically, a well-shot flat lay on a neutral surface is enough for AI to generate ghost mannequin or on-model results, but the garment must be styled first — steamed, seams straight, sleeves positioned deliberately. AI treats an unstyled garment as an accurate record of a rumpled product.
A pre-shoot checklist you can hand to anyone
The value of writing this down is that source capture is delegable. A warehouse associate with a phone and this list produces better AI input than a photographer improvising without it.
| Step | Check |
|---|---|
| Product prep | Cleaned, dust removed, tags tucked, garments steamed |
| Surface | Plain, mid-tone, non-reflective; no patterned tabletops |
| Light | Single source, diffused, room lights off, fill card opposite |
| Camera | Highest resolution mode, HDR off, flash off |
| Focus | Locked on the product's key detail, verified at 100% zoom |
| Exposure | No clipped whites; check the histogram or highlight warning |
| Framing | Whole product visible with margin on all four sides |
| Stability | Tripod or braced; no handheld below 1/60s |
| Coverage | Minimum 5 angles per SKU before moving on |
| Naming | SKU-angle convention applied at capture, not later |
That last row saves more time than it appears to. When you batch dozens of SKUs through an AI pipeline, the bottleneck is almost never generation — it is matching outputs back to the right product. Naming files SKU1234-front.jpg at capture makes the whole downstream process traceable.
Once your inputs meet this bar, the generation step becomes genuinely reliable rather than a coin flip. Retouchable's pipeline, for example, will happily produce ghost mannequin, on-model, and lifestyle variations from a single well-shot flat lay — but the same pipeline fed a blurry, clipped source will surface those flaws in every variation it produces. Garbage in is not a cliché here; it is a mathematical property of how these models work.
Shoot one SKU, run it through your full AI workflow, and inspect the output at 100%. Fix the capture problems you find, then shoot the remaining 200 products. Reversing that order is the most expensive mistake in catalog photography.