Original-versus-latest AI image editing test shown as one source photo branching into two edited sequences

Original Image or Latest Edit? What Happened in Our 12-Edit ChatGPT Test

by

You fix one detail in an AI image, then something else shifts. Should the next edit start from the untouched original or the latest result?

The tested takeaway

In our two synthetic examples, neither route came out ahead. The toolbox edits both worked but created different details; the person edits both let rain reflections spill across a floor that was supposed to stay dry.

What we tested: An AI agent ran three edits on each of two synthetic images. One workflow continued from the latest result; the other returned to the untouched original and repeated every requested change. That made 12 edits in total, with one run for each operation.

The choice: restart from the original or continue from the latest edit?

After a useful AI image edit, the next step creates a real tradeoff. Using the latest result gives the model those new details—and any mistakes—as its next reference. Returning to the original image restores the clean starting point, but the tool must generate every requested change again from the written instructions.

We asked a concrete question: after three cumulative edits, would either workflow preserve the scene better? The answer was a tie in both examples, but the pictures reveal more than the tie. One pair succeeded with visibly different details. The other pair repeated the same boundary mistake.

What the images show

Example 1 · Both routes completed the request

The toolbox edit worked both ways — with different details

Both final images delivered the requested green toolbox, red-handled screwdriver, and plywood worktop. Look closer, though, and the screwdriver design and wood grain differ. Starting again from the original recreated the requested idea, not the exact object generated one step earlier.

Synthetic equipment bench source with navy toolbox, orange multimeter, wrench, cable, and blank clipboard.
Original image: The untouched equipment scene before the toolbox, tool, and worktop changes.
Stage 3 latest-output arm output showing a forest-green toolbox and red-handled screwdriver on a warm plywood bench surface.
Continued from the latest result: All three requested changes are present. This version carried the generated screwdriver forward from the previous edit.
Stage 3 original-source arm output showing a forest-green toolbox and red-handled screwdriver on a warm plywood bench surface.
Restarted from the original image: All three requested changes are present, but the screwdriver and wood grain were generated again and look different.

What this means: If the exact screwdriver, texture, or arrangement from a previous edit matters, the latest result gives the next edit that specific detail to work from. If only the requested category matters, returning to the original may still produce a usable but visibly different version.

Example 2 · Both routes crossed the same boundary

The doorway edit failed at the same boundary both ways

Both final images added the teal jacket, hearing protectors, and rainy exterior. Both also changed the scene where they should not have: wet reflections crossed the doorway threshold and spread onto the interior floor. Minor facial drift appeared too.

Synthetic field project manager in an orange jacket holding a black tablet in an unfinished interior.
Original image: The untouched person scene before the jacket, hearing-protection, and weather changes.
Stage 3 latest-output arm output showing the fictional field manager in a teal jacket with yellow hearing protectors and a rainy doorway view.
Continued from the latest result: The rainy exterior is present, but wet reflections also spread across the threshold onto the interior floor.
Stage 3 original-source arm output showing the fictional field manager in a teal jacket with yellow hearing protectors and a rainy doorway view.
Restarted from the original image: Restarting from the original did not prevent the same spill onto the interior floor.

What this means: Returning to the original was not a reset button for this mistake. When an edit stops at a doorway, product edge, or other hard boundary, name the area that must stay unchanged and inspect that boundary at full size before continuing.

A practical checkpoint workflow for your next edit

Choose the reference that contains what matters most: start from the original when the untouched scene is the priority, or continue from the latest result when a newly created detail must carry forward. Save each good checkpoint and inspect the edge of every change at full size.

These are practical suggestions from our observations, not a separately tested workflow.

  1. Save the good version. Keep the original and every acceptable checkpoint before asking for another change.
  2. State the full target. Repeat every earlier change that must still appear, plus the new request.
  3. Name what must stay untouched. Be specific about the person, object, position, framing, lighting, and regions outside the edit.
  4. Choose your reference. Use the latest result when its new details matter; return to the original when the clean baseline matters more.
  5. Inspect the boundary. Zoom in around doorways, product edges, faces, hands, shadows, and reflections.
  6. Allow for cleanup. A convincing image can still miss a narrow instruction.

For a doorway like this one, a clearer follow-up could say: “Change only the area beyond the doorway threshold; keep the interior floor, reflections, and light pattern unchanged.” We did not test this follow-up.

See all 12 edits

The complete record is available below for readers who want to inspect every stage. The first edits used the same image and prompt. The workflows diverged from the second edit onward; the third edit was our main comparison.

Open the complete 12-edit record

Equipment bench

Edit 1 of 3

Continued from the latest result
Stage 1 latest-output arm output showing a forest-green toolbox on the synthetic equipment bench.
Equipment bench, edit 1: Latest-output arm, source A, stage 1: raw output showing a forest-green toolbox on the synthetic equipment bench; shown unaltered.
Prompt used
Equipment bench, cumulative edit 1
Initial hidden-workflow score
100.0/100
Audited score
100.0/100
Requested changes
The toolbox body is uniformly forest green while the lid panel, molded ribs, latches, handle, scale, perspective, highlights, and contact shadows remain visibly consistent with R0.
What was meant to stay unchanged
The three-quarter overhead camera angle, framing, crop, bench edges, and relative spacing are preserved. The orange multimeter, blank display, dial, probes, leads, scale, and left placement match R0; pre-existing dial markings remain baseline detail. The right-side coiled black cable, connectors, coil geometry, and placement are unchanged. The upper-right clipboard retains its blank white paper, clip, dark border, perspective, and position. Apart from the requested color, toolbox dimensions, molded form, centered placement, latches, handle, and hardware are preserved. Soft upper-left lighting, highlights, bench falloff, and all contact shadows remain coherent and source-matched.
Other visible changes
The wrench, cable, meter, clipboard, plant, and background storage context remain present; no new material object is visible.
Dimensions
1536 × 1024 pixels
Usability
The single requested recolor is clear, natural, and usable without visible cleanup.
Failure state
No generation failure recorded.

Observed: The single requested recolor is clear, natural, and usable without visible cleanup.

Restarted from the original image
Stage 1 original-source arm output showing a forest-green toolbox on the synthetic equipment bench.
Equipment bench, edit 1: Original-source arm, source A, stage 1: raw output showing a forest-green toolbox on the synthetic equipment bench; shown unaltered.
Prompt used
Equipment bench, cumulative edit 1
Initial hidden-workflow score
100.0/100
Audited score
100.0/100
Requested changes
The toolbox body is consistently forest green while its molded panels, handle, latches, scale, perspective, highlights, and contact shadows remain intact.
What was meant to stay unchanged
Three-quarter overhead camera, framing, crop, bench edges, and relative spacing are preserved. Orange multimeter, blank display, dial, probes, leads, scale, and left placement match R0. Right-side coiled cable, connectors, coil geometry, and placement remain unchanged. Upper-right clipboard preserves its blank paper, clip, black border, perspective, and position. Apart from requested color, toolbox dimensions, molded form, centered placement, latches, handle, and hardware match R0. Soft upper-left light, bench falloff, highlights, and all object contact shadows remain source-matched.
Other visible changes
Wrench, meter, cable, clipboard, plant, and background storage context remain; no new material object is visible.
Dimensions
1536 × 1024 pixels
Usability
The requested recolor is clear, natural, and usable without visible cleanup.
Failure state
No generation failure recorded.

Observed: The requested recolor is clear, natural, and usable without visible cleanup.

Edit 2 of 3

Continued from the latest result
Stage 2 latest-output arm output showing a forest-green toolbox and red-handled screwdriver on the synthetic equipment bench.
Equipment bench, edit 2: Latest-output arm, source A, stage 2: raw output showing a forest-green toolbox and red-handled screwdriver on the synthetic equipment bench; shown unaltered.
Prompt used
Equipment bench, cumulative edit 2
Initial hidden-workflow score
100.0/100
Audited score
100.0/100
Requested changes
The toolbox is consistently forest green while its size, molded lid, three latches, handle, perspective, highlights, and shadows remain intact. One realistic red-handled flat-head screwdriver occupies the former wrench's lower-left diagonal footprint and orientation; no wrench remains.
What was meant to stay unchanged
Three-quarter overhead camera, framing, crop, bench boundaries, and object layout match R0. Orange multimeter retains its blank display, dial, controls, probes, leads, scale, and left-side position. Coiled black cable remains on the right with matching coil geometry, connectors, scale, and orientation. White blank clipboard remains upper-right with matching clip, border, perspective, and position. Apart from color, toolbox dimensions, molded form, centered placement, handle, latches, and hardware remain source-faithful. Upper-left lighting direction, highlights, contact shadows, and overall bench illumination remain coherent and source-matched.
Other visible changes
No material addition/removal is visible beyond the requested wrench-to-screwdriver replacement.
Dimensions
1536 × 1024 pixels
Usability
Both edits are immediately legible and require no visible cleanup.
Failure state
No generation failure recorded.

Observed: Both edits are immediately legible and require no visible cleanup.

Restarted from the original image
Stage 2 original-source arm output showing a forest-green toolbox and red-handled screwdriver on the synthetic equipment bench.
Equipment bench, edit 2: Original-source arm, source A, stage 2: raw output showing a forest-green toolbox and red-handled screwdriver on the synthetic equipment bench; shown unaltered.
Prompt used
Equipment bench, cumulative edit 2
Initial hidden-workflow score
100.0/100
Audited score
100.0/100
Requested changes
The hard-shell body is a consistent dark forest green while its centered size, molded ribs, recessed lid panel, three latches, handle, perspective, highlights, and shadows remain intact. One realistic red translucent-handled flat-head screwdriver replaces the wrench, following the same lower-left-to-lower-center diagonal footprint and orientation; no wrench remains.
What was meant to stay unchanged
The three-quarter overhead angle, framing, crop, bench boundaries, and object layout match R0. The orange multimeter retains its blank display, dial, controls, probes, leads, scale, and left-side position. The coiled black cable remains at the right with the same coil count, plug placement, scale, and orientation. The white blank clipboard remains in the upper-right position with the same black edge, clip, perspective, and empty paper. Apart from the requested green color, toolbox size, centered placement, molded shape, handle, latches, and hardware visibly match R0. Upper-left window light, object highlights, contact shadows, and bench illumination remain coherent and visibly source-matched.
Other visible changes
No material addition or removal is visible beyond the requested wrench-to-screwdriver replacement.
Dimensions
1536 × 1024 pixels
Usability
Both requested changes are immediately legible and integrated without a visible cleanup need.
Failure state
No generation failure recorded.

Observed: Both requested changes are immediately legible and integrated without a visible cleanup need.

Edit 3 of 3

Continued from the latest result
Stage 3 latest-output arm output showing a forest-green toolbox and red-handled screwdriver on a warm plywood bench surface.
Equipment bench, edit 3: Latest-output arm, source A, stage 3: raw output showing a forest-green toolbox and red-handled screwdriver on a warm plywood bench surface; shown unaltered.
Prompt used
Equipment bench, cumulative edit 3
Initial hidden-workflow score
100.0/100
Audited score
100.0/100
Requested changes
The toolbox is consistently forest green while its molded lid, latches, handle, scale, perspective, highlights, and contact shadows remain intact. One realistic red-handled flat-head screwdriver follows the former wrench's lower-left diagonal footprint and orientation; no wrench remains. The bench top is warm light plywood with visible but coherent natural grain and knots, believable edge treatment, and integrated contact shadows.
What was meant to stay unchanged
Three-quarter overhead camera, framing, crop, bench boundaries, and relative object placement remain source-aligned. Orange multimeter, blank display, dial, probes, leads, scale, and left position are preserved. Right-side coiled cable retains its coil geometry, connectors, size, and orientation. Upper-right clipboard preserves blank paper, black border, clip, perspective, and position. Apart from requested color, toolbox dimensions, molded shape, centered placement, handle, latches, and hardware match R0. Upper-left light direction, window-shadow pattern, object highlights, and coherent contact shadows remain plausible on the new surface.
Other visible changes
No material addition/removal is visible beyond the requested tool replacement and surface change.
Dimensions
1536 × 1024 pixels
Usability
All cumulative changes are clear and integrated without a visible cleanup need.
Failure state
No generation failure recorded.

Observed: All cumulative changes are clear and integrated without a visible cleanup need.

Restarted from the original image
Stage 3 original-source arm output showing a forest-green toolbox and red-handled screwdriver on a warm plywood bench surface.
Equipment bench, edit 3: Original-source arm, source A, stage 3: raw output showing a forest-green toolbox and red-handled screwdriver on a warm plywood bench surface; shown unaltered.
Prompt used
Equipment bench, cumulative edit 3
Initial hidden-workflow score
100.0/100
Audited score
100.0/100
Requested changes
The toolbox is consistently forest green while its molded lid, three latches, handle, centered scale, perspective, highlights, and contact shadows remain intact. One realistic red-handled flat-head screwdriver follows the former wrench's lower-left diagonal footprint and orientation; no wrench remains. The full bench top is warm light plywood with restrained grain, believable edge treatment, and coherent object contact shadows.
What was meant to stay unchanged
Three-quarter overhead camera, framing, crop, bench boundaries, and relative object layout match R0. Orange multimeter, blank display, dial, probes, leads, scale, and left position are preserved. Right-side coiled black cable retains its coil geometry, connector placement, scale, and orientation. Upper-right clipboard retains blank paper, black edge, clip, perspective, and position. Apart from requested color, toolbox dimensions, molded form, centered placement, handle, latches, and hardware match R0. Upper-left light direction, highlights, window-shadow pattern, and coherent contact shadows are preserved on the new surface.
Other visible changes
No material addition or removal is visible beyond the requested wrench-to-screwdriver replacement and surface change.
Dimensions
1536 × 1024 pixels
Usability
All three cumulative edits are clear and integrated without a visible cleanup need.
Failure state
No generation failure recorded.

Observed: All three cumulative edits are clear and integrated without a visible cleanup need.

Fictional field manager

Edit 1 of 3

Continued from the latest result
Stage 1 latest-output arm output showing the fictional field manager in a teal jacket.
Fictional field manager, edit 1: Latest-output arm, source B, stage 1: raw output showing the fictional field manager in a teal jacket; shown unaltered.
Prompt used
Fictional field manager, cumulative edit 1
Initial hidden-workflow score
100.0/100
Audited score
100.0/100
Requested changes
The jacket is consistently teal across the torso and both sleeves; the black collar, zipper, seams, pocket, fabric folds, fit, and directional shading remain visibly coherent with R0.
What was meant to stay unchanged
Face shape, features, skin tone, short dark curls, beard detail, eye direction, and slight smile match R0 without visible identity drift. The waist-up pose, shoulder angle, right arm, visible hand, finger placement, and tablet grip remain in the same positions. The closed black tablet retains the same shape, orientation, size, edge highlights, and chest-height placement. The gray hard hat remains on the same right-side shelf with the same scale and orientation. The blue tape roll remains on the left workbench; surrounding paper rolls and bench items are unchanged. Framing, eye-level camera, doorway geometry, studs, shelving, daylight exterior, interior exposure, and subject lighting visibly match R0.
Other visible changes
No material object or person was visibly added or removed.
Dimensions
1536 × 1024 pixels
Usability
The requested recolor reads naturally at fit view and native detail and requires no visible cleanup for the stated fictional job.
Failure state
No generation failure recorded.

Observed: The requested recolor reads naturally at fit view and native detail and requires no visible cleanup for the stated fictional job.

Restarted from the original image
Stage 1 original-source arm output showing the fictional field manager in a teal jacket.
Fictional field manager, edit 1: Original-source arm, source B, stage 1: raw output showing the fictional field manager in a teal jacket; shown unaltered.
Prompt used
Fictional field manager, cumulative edit 1
Initial hidden-workflow score
100.0/100
Audited score
100.0/100
Requested changes
The jacket is uniformly teal across torso and sleeves while its black collar, zipper, seams, pocket, fit, folds, and shading remain coherent.
What was meant to stay unchanged
Face, skin tone, hair, apparent identity, beard detail, eye direction, and slight smile remain visibly source-faithful. Body pose, shoulder angle, arms, visible hand, fingers, and tablet grip remain in the same positions. Closed black tablet retains its shape, size, orientation, highlights, and chest-height placement. Gray hard hat remains unchanged on the right-side shelf. Blue tape roll and left workbench context remain source-matched. Camera, framing, crop, doorway geometry, studs, shelves, daylight exterior, indoor exposure, and subject lighting match R0.
Other visible changes
No material object or person was visibly added or removed.
Dimensions
1536 × 1024 pixels
Usability
The single requested recolor is natural and usable without visible cleanup.
Failure state
No generation failure recorded.

Observed: The single requested recolor is natural and usable without visible cleanup.

Edit 2 of 3

Continued from the latest result
Stage 2 latest-output arm output showing the fictional field manager in a teal jacket with yellow hearing protectors.
Fictional field manager, edit 2: Latest-output arm, source B, stage 2: raw output showing the fictional field manager in a teal jacket with yellow hearing protectors; shown unaltered.
Prompt used
Fictional field manager, cumulative edit 2
Initial hidden-workflow score
97.5/100
Audited score
97.5/100
Requested changes
The jacket is consistently teal with its black collar, zipper, seams, pocket, cut, fit, folds, and directional shadows coherent. One plausible yellow-and-black protector set rests around the neck while leaving the face, hair, hand, tablet, and jacket closure visible.
What was meant to stay unchanged
The person remains recognizably the same with matching skin tone and hair, but eyes, nose, mouth, cheek contours, and beard detail show minor visible regeneration drift. Waist-up pose, shoulders, arms, visible hand, fingers, and tablet grip remain aligned with R0. The closed black tablet retains its shape, size, orientation, edge highlights, and chest-height placement. The gray hard hat remains unchanged on the right shelf. The blue tape roll and left workbench papers and rolls remain source-matched. Camera, framing, crop, room geometry, clear daylight doorway, indoor exposure, and subject lighting remain aligned with R0.
Other visible changes
No material object or person was added or removed beyond the requested hearing protectors.
Dimensions
1536 × 1024 pixels
Usability
The output remains usable as delivered; the facial drift is minor for the fictional use case.
Failure state
No generation failure recorded.

Observed: The jacket and protectors were delivered, but the reviewer found minor facial regeneration drift.

Restarted from the original image
Stage 2 original-source arm output showing the fictional field manager in a teal jacket with yellow hearing protectors.
Fictional field manager, edit 2: Original-source arm, source B, stage 2: raw output showing the fictional field manager in a teal jacket with yellow hearing protectors; shown unaltered.
Prompt used
Fictional field manager, cumulative edit 2
Initial hidden-workflow score
100.0/100
Audited score
100.0/100
Requested changes
The jacket is uniformly teal while its black collar, zipper, seams, chest pocket, cut, folds, fit, and shadows remain coherent. One realistic yellow-and-black protector set rests around the neck without covering the face, hair, visible hand, tablet, or central zipper closure.
What was meant to stay unchanged
Face shape, features, skin tone, short dark curls, beard detail, eye direction, and expression remain visibly source-faithful. Pose, shoulders, arms, visible hand, fingers, and tablet grip remain in their source positions. Black tablet keeps the same shape, orientation, scale, and chest-height placement. Gray hard hat remains unchanged on the right shelf. Blue tape roll and left workbench context remain source-matched. Camera, framing, crop, room geometry, daylight doorway, interior exposure, and subject lighting match R0.
Other visible changes
No material object or person was added or removed beyond the requested hearing protectors.
Dimensions
1536 × 1024 pixels
Usability
Both requested edits are clear and usable without visible cleanup.
Failure state
No generation failure recorded.

Observed: Both requested edits are clear and usable without visible cleanup.

Edit 3 of 3

Continued from the latest result
Stage 3 latest-output arm output showing the fictional field manager in a teal jacket with yellow hearing protectors and a rainy doorway view.
Fictional field manager, edit 3: Latest-output arm, source B, stage 3: raw output showing the fictional field manager in a teal jacket with yellow hearing protectors and a rainy doorway view; shown unaltered.
Prompt used
Fictional field manager, cumulative edit 3
Initial hidden-workflow score
97.5/100
Audited score
78.3/100
Requested changes
The jacket is consistently teal across torso and sleeves while its black collar, zipper, seams, pocket, fit, folds, and shading remain coherent. One plausible yellow-and-black over-ear protector set rests naturally around the neck without covering the face, hair, hand, tablet, or central jacket closure. Visible rainfall, gray overcast light, softened distance, and wet reflective exterior ground replace the clear-day doorway view while the indoor scene remains readable.
What was meant to stay unchanged
Skin tone, hair, and apparent identity remain recognizable, but the eyes, nose, mouth, cheek contours, and beard texture show minor visible regeneration drift from R0. Waist-up pose, shoulder angle, arms, visible hand, finger placement, and tablet grip remain source-matched. Closed black tablet retains the same shape, size, orientation, edge highlights, and chest-height placement. Gray hard hat remains on the same right-side shelf with matching scale and orientation. Blue tape roll and surrounding left-workbench papers and rolls remain in their source positions. Camera, crop, studs, shelving, doorway structure, and indoor spatial geometry remain aligned; indoor exposure stays believable despite the requested exterior weather change. The later full-size review found that the interior floor changed across the doorway threshold.
Other visible changes
Post-lock audit: wet reflective treatment extended from the requested exterior into the protected interior floor behind the workbench.
Dimensions
1536 × 1024 pixels
Usability
The result needs a localized edit or cleanup pass before it meets the exterior-only instruction.
Failure state
No generation failure recorded.

Observed: The cumulative changes are visible, but minor face drift remained and the later full-size review found wet reflections crossing the doorway threshold onto the interior floor. A localized edit or cleanup pass would be needed.

Restarted from the original image
Stage 3 original-source arm output showing the fictional field manager in a teal jacket with yellow hearing protectors and a rainy doorway view.
Fictional field manager, edit 3: Original-source arm, source B, stage 3: raw output showing the fictional field manager in a teal jacket with yellow hearing protectors and a rainy doorway view; shown unaltered.
Prompt used
Fictional field manager, cumulative edit 3
Initial hidden-workflow score
97.5/100
Audited score
78.3/100
Requested changes
Teal color is consistent across the jacket while its black collar, zipper, seams, pocket, fit, folds, and shadows remain coherent. One plausible yellow-and-black protector set rests naturally around the neck and leaves the face, hair, hand, tablet, and jacket closure visible. Rain streaks, gray overcast sky, softened exterior detail, and wet reflective ground are visible through the doorway while the indoor subject remains believably lit.
What was meant to stay unchanged
Apparent identity, skin tone, and hair remain recognizable, but eye shape, nose, mouth, cheek modeling, and beard texture show minor visible drift from R0. Body pose, shoulders, arms, visible hand, finger placement, and tablet grip remain source-matched. Closed black tablet retains its shape, size, orientation, highlights, and chest-height position. Gray hard hat remains on the same right-side shelf. Blue tape roll and left workbench context remain in the same locations. Camera angle, crop, doorway frame, studs, shelving, room geometry, and interior exposure remain source-aligned around the requested exterior change. The later full-size review found that the interior floor changed across the doorway threshold.
Other visible changes
Post-lock audit: wet reflective treatment extended from the requested exterior into the protected interior floor behind the workbench.
Dimensions
1536 × 1024 pixels
Usability
The result needs a localized edit or cleanup pass before it meets the exterior-only instruction.
Failure state
No generation failure recorded.

Observed: The cumulative changes are visible, but minor face drift remained and the later full-size review found wet reflections crossing the doorway threshold onto the interior floor. A localized edit or cleanup pass would be needed.

How the comparison was run

Each source received the same three cumulative instructions. For one workflow, each new edit used the previous result. For the other, every edit started from the untouched source. At the same stage, both workflows received exactly the same written prompt.

A TPTK AI agent operated the experiment through Si Fisher’s authorized ChatGPT account; Si did not personally run or score the edits. Each scored edit used a fresh ChatGPT Images entry with Instant selected and no deliberate prior-chat context.

Scoring, review, and disclosure details

An AI reviewer first scored the raw files without knowing which workflow produced them. A separate agent then inspected every source and result with the workflow labels visible. A focused second review checked the doorway issue after it was noticed.

The toolbox finals each scored 100/100. That means they received full credit under the test’s visual rubric; it does not mean the images were identical. The person finals initially scored 97.5/100 for minor facial drift. After the doorway spill was reviewed, the accepted audited score was 78.3/100 for each, while the original 97.5 records were preserved.

The review considered whether requested changes appeared, whether six specified scene details stayed intact, whether anything was added or removed without instruction, visible artifacts, plausibility, and practical usability. The images and concrete observations are more useful than the summary numbers.

An earlier Temporary-chat attempt reported that image editing was unavailable and produced no image. It is preserved as an access record and was not scored.

What this test cannot establish

  • Two synthetic sources and one run per condition cannot estimate general reliability or represent real client photographs.
  • The visible consumer surface label does not identify the hidden backend model.
  • Fresh regular entries exclude deliberately supplied chat history; account memory, session independence, personalization, routing, seed, and hidden platform context were not exposed.
  • The source prompts intentionally requested no readable copy or logos, but incidental preexisting meter marks and background papers remained frozen baseline details.
  • The two workflows tied in these examples. That does not mean they will perform equally on other images or instructions.

For another TPTK test built around frozen inputs and preserved outputs, read the ChatGPT vs Gemini real-estate flyer test.

Exact cumulative prompts

These are the exact prompts used in the frozen comparison. They demonstrate the structure of a cumulative target and preservation list; copying them does not guarantee the same output.

Download all six exact prompts

Equipment bench — stage 1 
Edit only the attached source image. Treat the instructions below as the complete target for this output, not as a reference to any earlier conversation.

Complete cumulative target:
1. Change the closed hard-shell toolbox body from navy blue to forest green. Preserve its size, shape, centered position, latches, handle, perspective, and shadows.

Preserve exactly: the locked camera angle, framing, crop, orange multimeter with blank display on the left, coiled black cable on the right, silver adjustable wrench in the lower-left area, white clipboard with blank paper in the upper-right area, pale gray concrete work surface, upper-left lighting direction, and all unmentioned object positions.

Make no other changes. Add no people, hands, text, numbers, logos, labels, prices, interfaces, brand marks, or new objects. Return exactly one image at the source aspect ratio.
Equipment bench — stage 2 
Edit only the attached source image. Treat the instructions below as the complete target for this output, not as a reference to any earlier conversation.

Complete cumulative target:
1. The closed hard-shell toolbox body is forest green. Preserve its size, shape, centered position, latches, handle, perspective, and shadows.
2. Replace the silver adjustable wrench in the lower-left area with one realistic red-handled flat-head screwdriver occupying the same general footprint and orientation.

Preserve exactly: the locked camera angle, framing, crop, orange multimeter with blank display on the left, coiled black cable on the right, white clipboard with blank paper in the upper-right area, pale gray concrete work surface, upper-left lighting direction, and all unmentioned object positions.

Make no other changes. Add no people, hands, text, numbers, logos, labels, prices, interfaces, brand marks, or extra objects. Return exactly one image at the source aspect ratio.
Equipment bench — stage 3 
Edit only the attached source image. Treat the instructions below as the complete target for this output, not as a reference to any earlier conversation.

Complete cumulative target:
1. The closed hard-shell toolbox body is forest green. Preserve its size, shape, centered position, latches, handle, perspective, and shadows.
2. A realistic red-handled flat-head screwdriver occupies the former wrench's lower-left footprint and orientation; no wrench remains.
3. Change only the bench surface material from pale gray concrete to warm light plywood with subtle realistic grain. Preserve the existing object placement, camera, lighting direction, and coherent contact shadows.

Preserve exactly: the locked camera angle, framing, crop, orange multimeter with blank display on the left, coiled black cable on the right, white clipboard with blank paper in the upper-right area, upper-left lighting direction, and all unmentioned object positions.

Make no other changes. Add no people, hands, text, numbers, logos, labels, prices, interfaces, brand marks, or extra objects. Return exactly one image at the source aspect ratio.
Field manager — stage 1 
Edit only the attached source image. Treat the instructions below as the complete target for this output, not as a reference to any earlier conversation.

Complete cumulative target:
1. Change the fictional field project manager's rust-orange work jacket to teal. Preserve the jacket's cut, texture, fit, folds, and shadows.

Preserve exactly: the person's face, skin tone, hair, apparent identity, expression, body pose, hands, closed black tablet and its position, gray hard hat on the shelf, blue tape roll on the left workbench, camera angle, framing, crop, room geometry, daylight doorway, and indoor lighting.

Make no other changes. Add no readable text, logos, labels, interfaces, brand marks, or other people. Return exactly one image at the source aspect ratio.
Field manager — stage 2 
Edit only the attached source image. Treat the instructions below as the complete target for this output, not as a reference to any earlier conversation.

Complete cumulative target:
1. The fictional field project manager's work jacket is teal. Preserve the jacket's cut, texture, fit, folds, and shadows.
2. Add one pair of realistic yellow over-ear hearing protectors resting around the person's neck. They must integrate naturally without covering the face, hair, hands, tablet, or jacket closure.

Preserve exactly: the person's face, skin tone, hair, apparent identity, expression, body pose, hands, closed black tablet and its position, gray hard hat on the shelf, blue tape roll on the left workbench, camera angle, framing, crop, room geometry, daylight doorway, and indoor lighting.

Make no other changes. Add no readable text, logos, labels, interfaces, brand marks, or other people. Return exactly one image at the source aspect ratio.
Field manager — stage 3 
Edit only the attached source image. Treat the instructions below as the complete target for this output, not as a reference to any earlier conversation.

Complete cumulative target:
1. The fictional field project manager's work jacket is teal. Preserve the jacket's cut, texture, fit, folds, and shadows.
2. One pair of realistic yellow over-ear hearing protectors rests naturally around the person's neck without covering the face, hair, hands, tablet, or jacket closure.
3. Change only the view through the open background doorway from clear daylight to visibly rainy overcast weather, including a softly wet exterior surface. Keep the interior exposure believable and preserve the room lighting and subject lighting.

Preserve exactly: the person's face, skin tone, hair, apparent identity, expression, body pose, hands, closed black tablet and its position, gray hard hat on the shelf, blue tape roll on the left workbench, camera angle, framing, crop, room geometry, and all unmentioned indoor objects.

Make no other changes. Add no readable text, logos, labels, interfaces, brand marks, or other people. Return exactly one image at the source aspect ratio.

Frequently asked questions

Is it better to edit the original image each time?

Not automatically. In these two examples, returning to the original recreated the requested toolbox changes with different visual details and did not prevent the shared doorway-boundary mistake.

When should I edit the latest output instead?

Use the latest result when the next instruction depends on a specific object, texture, layout, or design created in the previous edit. Save checkpoints so you can return to an earlier acceptable state.

Does a 100 score mean the images were identical?

No. It means the output received full credit under the visual scoring method used for this test. The two equipment finals visibly differed in screwdriver design and wood grain.

Did ChatGPT preserve the person perfectly?

No. The reviewer found minor facial regeneration drift, and both final branches changed the protected interior floor near the doorway. Both finals needed localized cleanup to meet the exterior-only instruction.

Can this test predict every ChatGPT image edit?

No. It used two synthetic sources, three stages, and one run per operation through the visible ChatGPT Images surface with Instant selected. Account context, seed, hidden routing, and backend model were not exposed.

Sources and evidence

Leave a Comment