PRACTICAL GUIDE
Paint over a small distraction, run the repair and inspect the surrounding texture. Repeated patterns and structural lines can expose the edit.
RECORDED EXAMPLE · OCTOBER 12, 2026
One short painted stroke removes the small glare on the left tower. Check the softened tower detail before accepting the repair; the earlier spoon failure is retained below.
SMALL-REPAIR TEST · OCTOBER 12, 2026
This new run repeats a narrow, achievable task: remove the small bright flare on the far-left tower, around source pixel (81,350). It does not remove the larger lamps along the bottom or the tower itself. We used the unchanged SpaceX photograph, a brush radius of 15 and one short stroke over the flare, followed by one processing pass.
The star-shaped light disappears, while the rocket, larger lights and main tower silhouette remain. The replacement is darker and softer around the thin tower member. At ordinary display size it removes the distraction; at full size the local texture is still synthetic. This is acceptable for a disclosed decorative edit, not evidence of what was hidden behind the glare. The older incomplete spoon mask below remains a failure example.
The record retains brush size, drag endpoints and canvas geometry. It is an approximate manual recipe, not an exported pixel mask: pointer sampling and resampling can change the exact result. The original JPEG carries an Adobe RGB profile; the PNG export has no embedded profile, so raw channel differences alone are not a repair-quality score.
Source: SpaceX Falcon 9/DSCOVR photograph, public domain, distributed by scikit-image. This is an illustrative alteration, not an unedited documentary photograph.
Observed in the Codex in-app browser on macOS arm64 using the local release preview. The displayed PNG data was saved byte-for-byte without image re-encoding. The automation's download-event wait timed out, so this artifact alone is not evidence that a normal browser download completed. See the run record and file dimensions and SHA-256 hashes for the exact scope.
OBJECT REMOVAL LAB · SEPTEMBER 7, 2026
These two new runs test different editing decisions. The coffee photograph deliberately uses an incomplete spoon mask; the rocket photograph uses a short stroke over one small glare. Both completed on the production Object Remover and exported PNGs. Completion is separate from acceptance: the first still fails its brief, while the second needs close inspection where the glare crossed a thin structure.
The runs used the in-app browser on macOS arm64 during approximately 13:38–13:46 China Standard Time. That is an observation window, not an inference-time benchmark. The recovered MI-GAN worker/model were in use, before the newer progress-status patch. Model and worker fingerprints are in the run record.
The intended delivery was a spoon-free coffee image. With brush radius 30, two strokes covered the bowl region but omitted much of the handle and related reflections. The exported bowl becomes darker and mottled; its metal rim and spoon-shaped silhouette remain readable. The handle above and alongside the cup still identifies the object. Bright reflected shapes on the cup also survive.
This is a useful failure because the operation returned an image without fulfilling the task. A smaller bright highlight is not equivalent to a missing spoon. For another attempt, I would trace the entire visible spoon, inspect its reflection separately, and check that the replacement follows the saucer's curvature and glossy texture. That is a proposed next edit, not a result established by these two strokes.
The target was the small star-shaped flare on the outer left tower, around pixel (81,350), not the larger lamps along the bottom. A radius-15 stroke ran from (81,347) to (82,354). An earlier misplaced mask was undone; only the final calibrated operation's saved export is shown.
The glare disappears, while the tower's main line, the central rocket and the other lamps remain recognizable. At the repaired point, the replacement has lower contrast. Follow the thin vertical member through that patch and compare the nearby lattice: removing a dazzling spot can expose an unconvincing continuation. This is usable as a small decorative repair at display size; the source cannot establish what the obscured tower detail actually looked like.
For coffee, the recorded screen drags were (422,440)→(428,492) and (409,453)→(424,483). Its displayed canvas began at (85,204), measured 562.5×375 CSS pixels and had a 1125×750 backing store. Mapping to the 600×400 source gives approximately (359.47,251.73)→(365.87,307.20) and (345.60,265.60)→(361.60,297.60).
The full record also preserves the rocket canvas geometry. We did not export either pixel mask or every intermediate pointer sample. Drag endpoints and brush settings support an approximate manual rerun; they do not guarantee identical antialiasing, mask resampling or output bytes.
The rocket JPEG embeds Adobe RGB (1998), while its exported PNG has no ICC profile. Comparing their raw channel values flags 106,601 pixels at a difference threshold of 16. Converting the source to sRGB in memory first reduces that count to 517, concentrated around the repaired light. The published image files were never transformed for this measurement.
| Case | Pixels differing by ≥16 in any RGB channel | Difference bounds [x0,y0,x1,y1) |
|---|---|---|
| Coffee · 600×400 | 5,579 | 317, 222, 396, 336 |
| Rocket · 640×427 | 517 | 67, 333, 96, 369 |
These counts locate changes; they are neither mask area nor reconstruction accuracy. Browser/Pillow decoding and color conversion can also differ. Download the four images and measurement script into one folder, install Pillow and NumPy, then run python3 measure-artifacts.py. It writes JSON measurements and hashes, never images.
Full measurements and color assumptions · SHA-256 manifest. To repeat the editing, upload the original file, apply the recorded brush operation, run once and save before retouching. Judge the whole object and adjacent structure before accepting the export.
Sources: Rachel Michetti's coffee photograph, courtesy of Pikolo Espresso Bar, CC0; SpaceX's Falcon 9/DSCOVR photograph, public domain, distributed by scikit-image. Source provenance and registry hashes. New material consists of the recorded edits, unchanged exports, measurements and review.
Cover the complete unwanted object, including shadows or reflections that belong to it, but avoid masking unrelated edges. A slightly expanded mask gives the model enough room to blend. Very broad masks force it to invent more of the scene and usually reduce reliability.
Remove isolated objects first and review each result before continuing. Walls, sky and soft ground textures are predictable; faces, hands, text, architecture and repeated patterns are more difficult. Several focused edits are often safer than one enormous selection.
Zoom out to judge whether the replacement fits the scene, then inspect at full resolution for repeated patches, blurred seams and bent lines. Compare the direction of floorboards, tiles, shadows and highlights with surrounding areas.
If an edit looks artificial, undo it rather than repeatedly processing the same damaged result. Reduce the mask, begin from a cleaner boundary and remove the object in logical pieces. A replacement often fails because the selected region contains too little surrounding evidence, crosses a unique line or asks the model to reconstruct hidden content that cannot be inferred. In those cases, a manual clone or crop may be more honest and reliable.
Object removal can improve composition, but it can also change the meaning of documentary or news imagery. Keep an original file, disclose material edits where accuracy matters and never use the tool to misrepresent evidence or remove ownership marks without permission.
REFERENCE
The model may not have enough nearby structure to reconstruct a convincing texture.
Yes, when the shadow clearly belongs to the removed object.
It may remove text, but rebuilding structured backgrounds behind it can be difficult.
Always keep an untouched source for comparison and future edits.
OBSERVED OUTPUT / OBJECT REMOVAL
The before image marks a pedestrian on the right side of the street scene in red. In the after image, that marked figure is absent and the surrounding storefront and pavement are reconstructed. The pair is useful precisely because the repaired area can be checked against nearby lines and texture.
EVIDENCE RECORD
REPRODUCE THE CHECK
The marked pedestrian is removed while the foreground portrait remains substantially unchanged. The edit reconstructs enough of the sidewalk and storefront to make the distraction less noticeable at ordinary viewing size.
The pair does not prove what was truly behind the pedestrian; that region is generated from surrounding context. Small differences in dimensions and compression also prevent a pixel-for-pixel quality claim. The result should not be described as recovery of hidden evidence.
Follow the storefront edge into the repaired area and out the other side. Repeat for curb direction and pavement texture, then compare blur and brightness across the boundary at normal size and 100 percent. Removing the pedestrian is only acceptable if the fill also fits its surroundings.
Repeated bricks, doubled poles, a locally smoother patch or a sudden change in JPEG texture can expose the fill. If one large mask produces these symptoms, divide the repair into smaller passes and preserve more surrounding context between runs.
