Hedra AI All articles
Experiments

Pixel Borders, Real Budgets: Testing Ten Image Editors Where the Mask Ends

Hedra AI
Pixel Borders, Real Budgets: Testing Ten Image Editors Where the Mask Ends

Photo: Giles Laurent, CC BY-SA 4.0, via Wikimedia Commons

At Hedra AI, we have a soft spot for ugly tests. Pretty demo images are easy. A clean seam on a real product photo, with a client waiting and a catalog deadline on the calendar, is harder. So we spent a week pushing the same masked edits through ten hosted image-editing APIs and staring at boundaries until our eyes crossed.

The brief was deliberately boring: remove a small logo from a textured surface, swap a studio backdrop for a neutral lifestyle scene, and repair a contact shadow under a product. Same source photos. Same masks, where the APIs allowed it. Same prompts, where prompts applied. Then we checked whether the edited pixels looked plausible and whether the unedited pixels survived untouched.

That second part is where things got interesting.

Why the seam is the whole product

Inpainting demos usually show the middle of the patch. Users do not live in the middle of the patch. They live at the edge, where generated texture meets original pixels. A perfect synthetic label is worthless if the surrounding paper grain suddenly goes smooth. A convincing background swap falls apart if the product casts a shadow in the wrong direction or picks up a halo the color of the old wall.

The seam is also where trust shows up. If an API quietly modifies pixels outside the mask, your retoucher has to re-check work they thought was done. If it repaints a logo into a slightly different logo, your legal team has a new problem. Boundary behavior is not a cosmetic detail. It is the contract between the model and the rest of your pipeline.

The test bench

We used ten hosted endpoints that could accept an image, a mask, and a text instruction. We labeled them Editor A through Editor J because the point is the protocol, not a leaderboard we want to defend forever. APIs change. Your mileage will vary by model version, image type, and whether the endpoint had coffee.

Our source set had twelve product photos: three matte cardboard boxes, three glossy bottles, three fabric items, and three brushed-metal objects. Each had at least one straight edge, one curved edge, and one textured region near the edit area. We made three edit types per image and two mask styles, a hard mask and a feathered mask, then ran three repeats where the API allowed it. That gave us a few hundred outputs to inspect, which is enough to find patterns and enough to make you hate looking at shadows.

The seam-audit protocol

Here is the repeatable version we would hand to a team choosing an editor in 2026.

  1. Freeze the source. Use PNG, not a re-compressed JPEG. Record dimensions, color profile, and checksum. If the API returns a different size, that is a finding, not a rounding error.

  2. Build three masks per edit. One tight hard mask, one feathered mask, and one deliberately sloppy mask. The sloppy mask tells you how the model handles ambiguity. Production masks are rarely perfect.

  3. Keep the prompt dull. We used instructions like remove the logo, replace the background with a neutral gray studio sweep, and repair the shadow beneath the object. Poetry belongs in the campaign, not the audit.

  4. Diff the unmasked pixels. Compare the output against the input outside the mask with a per-pixel color threshold. A few pixels of compression drift are normal. A repainted background outside the mask is a red flag. Count the percentage of changed pixels and inspect the worst offenders.

  5. Grade a five-pixel seam ring. Look for halos, doubled edges, color bleeding, texture stretching, and grain mismatch. Then zoom out. If the seam only works at 400 percent, it does not work.

  6. Score texture continuity separately. Grain, weave, brushed metal, and label typography all fail differently. A model can be great at flat walls and terrible at knit fabric.

  7. Run repeats. Same input, same mask, same prompt. If the seam moves by more than a couple of pixels or the shadow direction flips, your reviewer cannot build a reliable approval step around it.

What the ten editors did

We saw four rough clusters.

Hard-mask specialists respected the boundary and produced clean edges, but some were timid inside the mask. They left ghosting where a logo used to be or refused to invent a shadow. These are good candidates for catalog work where preserving the original product is more important than creative flourish.

Context painters were the opposite. They produced lovely backgrounds and convincing shadows, but several treated the mask as a suggestion. Unmasked pixels moved, product edges softened, and one API decided the best way to remove a logo was to redraw the entire label. Great for moodboards. Risky for SKU-level production.

Blenders split the difference. They feathered well and handled texture continuity better than the hard-mask crowd, but their seams depended heavily on mask quality. Give them a sloppy mask and they smeared the product edge into the background like a bad Photoshop clone-stamp session.

Pixel vandals were rare but memorable. One endpoint changed the color temperature of the whole image. Another added a subtle vignette nobody asked for. A third returned a lower-resolution image and called it a day. These are not seam problems. They are workflow problems wearing a seam costume.

A practical scorecard

If you are picking an editor for a US production pipeline, do not average everything into one magic number. Weight the things that break your review process. Our working scorecard looks like this: boundary integrity, 30 percent; untouched-pixel integrity, 25 percent; texture continuity, 20 percent; repeatability, 15 percent; operational friction, 10 percent.

Boundary integrity covers halos, doubled edges, and color bleed. Untouched-pixel integrity is the diff outside the mask. Texture continuity is grain, weave, and material response. Repeatability is how often two runs look like siblings instead of strangers. Operational friction is the unglamorous stuff: mask format, output size, latency, cost per edit, and whether the API returns anything useful when it fails.

For logo removal on textured packaging, we would weight untouched-pixel integrity higher. For background swaps in lifestyle scenes, boundary and shadow continuity matter more. For shadow repair under glossy products, repeatability climbs because a shadow that changes direction between runs will wreck a batch review.

Production notes before you commit

Ask for mask semantics in writing. Some APIs treat white as keep and black as edit. Some do the reverse. Some feather automatically. Some ignore your feather and apply their own. A five-minute documentation check saves an afternoon of inverted masks.

Test at your actual output size. A seam that looks fine at 1024 pixels can turn into a mushy border at 4K, especially on fabric and brushed metal. Also test your real file types. The difference between PNG and a twice-compressed JPEG is not academic when you are auditing boundaries.

Finally, keep a human in the loop for the first hundred edits. The scorecard narrows the field, but seams are judgment calls. A retoucher will spot the halo your pixel diff missed and the shadow your texture metric loved. That is not a failure of automation. It is the last mile of a production workflow that still depends on taste.

The bottom line

No editor won every category, which is exactly what we expected. The useful result is the protocol: fixed sources, multiple masks, unmasked-pixel diffs, seam-ring grading, texture checks, and repeat runs. If an image-editing API cannot survive that, it is not ready for your catalog. If it can, the seam stops being a surprise and starts being a spec.

All Articles

Related Articles

Your TTS Sounds Great Until the Feed Normalizes It: A Twelve-API Loudness Audit

Your TTS Sounds Great Until the Feed Normalizes It: A Twelve-API Loudness Audit

Aspect Ratio Is a Contract: Auditing Ten Image APIs for True Vertical Output

Aspect Ratio Is a Contract: Auditing Ten Image APIs for True Vertical Output

Seven Ways to Say No: What Seven Media APIs Actually Do When Your Prompt Trips a Filter

Seven Ways to Say No: What Seven Media APIs Actually Do When Your Prompt Trips a Filter