How to Edit Product Photos With AI Without Changing Them
A practical workflow for cleaning up product photos with AI while keeping the shape, label, color and proportions accurate.
Wan 2.6 Editorial
试用文中工具
Most AI photo editing is judged on how good the result looks. Product photography is judged on whether the result is still the same product. That flips the entire workflow. A background swap that makes a bottle look stylish but warps the cap is a failed edit. A relight that flatters a jacket but shifts the dye lot from navy to teal will get rejected by the listing owner, the marketplace, or the customer who files a return. The goal is not "make it prettier." The goal is "make it cleaner without making it different."
This guide is a working procedure for keeping the product intact through an AI edit: what to protect, what to photograph, how to prompt, how to review, and when to stop and reshoot. It pairs well with our AI product photo editor, and the prompt logic here builds on the scene-separation ideas in our text-to-image prompt structure guide.
试试这个配方
AI 产品照片编辑器
Why product accuracy is the hard part
Diffusion-style models are trained to produce plausible images, not faithful ones. When the input is ambiguous — a soft label, a dark seam, a reflective curve — the model fills the gap with something that looks reasonable. That is exactly what you want for a fantasy landscape and exactly what you do not want for a garment tag. The model has no concept of "this is the SKU." It only knows what it can resolve from the pixels and what your prompt implies.
Accuracy problems usually cluster in three places. First, fine repetitive detail: small text, serial numbers, stitching, mesh patterns. These get re-synthesized and drift. Second, identity-defining shape: a slightly tapered handle, an asymmetrical collar, the exact profile of a bottle shoulder. Mild warping reads as "wrong" to anyone who knows the product. Third, color under new lighting. Change the light, and the model may also change the material's apparent color, which matters when the catalog needs consistent hue across shots.
The practical consequence: treat every edit as a constraint problem. Decide what the model is allowed to touch before you run it, and review against that list afterward. If you cannot articulate which pixels are sacred, you are not ready to edit yet.
The four things the model must not change
Before editing, write down the non-negotiables for the specific product. Almost every product photo has the same four categories, and they are the review criteria later.
| Category | What it covers | Typical failure |
|---|---|---|
| Label text | Brand name, model number, ingredients, size, certifications | Letters re-rendered as plausible-looking gibberish; fonts drift |
| Silhouette | Outline, seams, proportions, hardware, closures | Handles soften, corners round, thickness changes |
| Color | Base hue, finish (matte/gloss), accents | Relighting shifts hue or adds a tint the material does not have |
| Proportions | Relative size of parts, symmetry where intended | A cap grows, a sleeve shortens, a logo shifts off-center |
For each item, note what the source photo actually shows. "Label is a black rectangle with white sans-serif text, three lines, small logo top-right." That sentence is your contract with the model and your checklist for review. If a category is already illegible in the source, it is not an edit problem — it is a capture problem, and no amount of prompting reliably reconstructs information that was never recorded.
A source photo checklist
Edits inherit the quality of the input. A sharp, evenly lit, well-framed source gives the model fewer gaps to invent. Run through this before uploading:
- Focus on the product, not the scene. Fine detail must be resolvable at 100% zoom. Blurry text cannot be faithfully preserved.
- Even, diffuse light on the product surface. Harsh specular highlights blow out label areas; deep shadows hide seams the model will guess at.
- A clean, uncluttered background. Busy backgrounds give the model more to confuse with the product edge.
- Full product in frame with breathing room. Cropped edges force the model to invent what was cut off.
- Consistent color reference. Include a neutral object or a known white/grey card in a test shot so you can sanity-check hue later.
- Multiple angles if the edit touches geometry. A single front shot does not tell the model what the side looks like; it will invent it.
- Original file, not a re-compressed screenshot. Compression artifacts read as texture and get amplified.
If the source fails on focus, label legibility, or framing, flag it now. You will save a generation or three by fixing it at the camera.
A prompt that separates the product from the scene
The reliable pattern is to describe the product as untouchable and describe only the environment, light, and framing as editable. Explicitly state what must be preserved, then describe the change. Vague prompts invite the model to reinterpret everything.
A short example for a matte black water bottle on a new studio background:
Product: a matte black stainless steel water bottle, 750ml, with a white printed label on the front and a brushed steel cap. Preserve the exact label text, the label position, the cap shape, and the bottle's tapered silhouette and proportions. Do not alter the product color or finish. Scene: place the bottle on a light grey seamless studio background with soft diffused key light from the left, gentle fill, and a subtle contact shadow. Keep the product sharply in focus. 50mm look, straight-on angle, product centered with even margins.
A few rules that make this work:
- Name the product twice — once to describe it, once in the preservation clause. Repetition reinforces the constraint.
- Separate product and scene with headings or line breaks. It reads as a boundary to the model and to you.
- Describe the light, not the mood. "Soft diffused key from the left" is actionable; "premium feel" is an invitation to restyle the product.
- State the angle. Changing the camera angle can silently change apparent proportions.
- Keep one variable per run. Change the background OR the light OR the crop. If you change all three and something breaks, you cannot tell which instruction caused it.
If you are new to structuring prompts, the scene-separation approach in the prompt structure guide applies directly here — same order, same discipline.
Reviewing the label and silhouette
Review at 100% zoom, side by side with the source, and check in a fixed order so nothing slips.
- Label text, character by character. Read it aloud. Compare letter shapes, not just words. Watch for substituted characters, dropped spaces, and font shifts.
- Silhouette overlay. Flick between source and output. Does the outline sit in the same place? Check handles, caps, corners, and any asymmetric feature.
- Proportions. Compare part-to-part ratios: cap to body, sleeve to shoulder, logo to edge. Small drifts compound across a catalog.
- Color. Sample the same spot in both images. A small hue shift is easy to miss and hard to excuse.
- Material surface. Matte should stay matte; gloss highlights should fall where the geometry says they should.
- Edges and contact shadow. Check for halos, smeared edges, or shadows that ignore the new light direction.
If any item fails, do not patch by stacking prompts on a bad result. Regenerate from the source with a tighter preservation clause. Stacked edits drift further each time. If a second attempt fails the same check, the source is likely the limit — move to the reshoot decision.
Keeping a catalog visually consistent
A single clean edit is a win; a consistent set is a catalog. Decide the shared parameters before you edit the first image, and write them down: background tone, light direction and softness, camera angle, framing margins, and shadow style. Apply the same prompt skeleton to every product, changing only the product description and its preservation clause.
Two habits keep sets honest. First, edit against a reference image from the approved set, not from memory — compare the new output to a known-good neighbor. Second, spot-check hue across the batch with a color sampler rather than your eye, since displays and fatigue both lie. If one product's color drifts, fix that one before adding more; drift spreads through a set faster than most people expect.
Keep an edit log per SKU: source file, prompt used, pass/fail notes. When a listing is questioned later, you can show exactly what changed and what was preserved. The workflow and cost per run are shown in the editor before you submit, so you can budget batches without guessing — see pricing for how credits work.
When a reshoot is cheaper than an edit
Editing is cheaper than a studio day only when the source already contains the information the output needs. It cannot recover what was never captured. Stop and reshoot when:
- The label is unreadable or blurred in the source. No prompt reliably reconstructs lost characters.
- The product is cropped or occluded. The model will invent the missing geometry, and it will be wrong.
- Glare or deep shadow destroys surface detail the edit needs to preserve.
- You need a new angle. Reposing a product from a single view is reconstruction, not editing.
- The product has complex, identity-critical texture — woven mesh, engraved patterns, dense print — that the model keeps smoothing.
- Repeated attempts fail the same preservation check. Two or three disciplined tries is the signal, not ten.
A useful rule: if the edit requires the model to know something the photo does not show, it is a capture problem. Twenty minutes with a light and a tripod usually beats an afternoon of regenerating and reviewing.
The through-line is simple. Edit the scene, protect the product, review against a written contract, and reshoot when the source cannot support the claim. That is how AI editing becomes a reliable part of a product workflow instead of a gamble you have to re-check every time.