
SINGLE-PRODUCT STUDY · IMAGE EDITING & RECOVERY
The Fix That Wouldn’t Stay Fixed
A Nautical Photo Shoot
How repeated image corrections became a test of identity preservation, repair stability, and the difference between acceptance and successful recovery.
A naturalistic evaluation of whether an AI image-generation workflow could correct visual defects while preserving my likeness, anatomy, clothing, and creative brief.
Role
Researcher · Evaluator · Annotation Designer
Completed
Full analysis and companion artifacts available below
Year
2026
Methods
Screenshot Analysis · Turn Annotation · Correction-Chain Tracing · Qualitative Visual Evaluation
Models/Products
ChatGPT with built-in image generation
Underlying image-model version not independently established
“A usable compromise is not the same as a verified repair.”
Suzanne Hansen
The image became acceptable. The repair remained unverified.
Correct the left foot while preserving the person, pose, outfit, and scene.
What I Requested
What the Record Supported
Repeated corrections left anatomical problems unresolved or introduced other defects. I eventually asked for different shoes to cover the toes.
An image with toe-covering flats. The result became usable, but a successful repair of the underlying foot anatomy was never demonstrated.
What I Accepted
ACCEPTED THROUGH WORKAROUND
→
At a Glance
A quick route into the finding, the record, and its limits
A corrected detail did not reliably produce an acceptable image, and an accepted image did not reliably preserve essential details through subsequent edits. Acceptance sometimes reflected a workaround or tolerated defect rather than complete repair.
The Finding
How I Checked
I traced 112 numbered turns through a 133-page annotated screenshot record, preserving my original outcome labels and following correction chains involving likeness, anatomy, clothing, scene geometry, and framing.
Evidence Boundary
This is one adaptive session with one user and one observed product workflow. Numbered turns can include multiple attempts or outputs. Turn outcomes do not establish an image-generation acceptance rate.
What the Record Shows
Thirteen turns were explicitly accepted, including three accepted through compromise. The notes also document 42 personal-photo attachments across ten explicitly counted reference-resupply events.
What to Inspect
Follow the foot-correction chain through repeated anatomical errors, my request to change shoes, and acceptance of the covered-foot result. Then compare acceptance with the separate question of whether the original defect was repaired.
The Investigation

The Question
Can an image-generation system reliably repair a personalized image while preserving identity, anatomy, and the creative brief, and what user effort does recovery require?

GENERATE → INSPECT → CORRECT → CHECK PRESERVATION → ACCEPT, RETRY, OR CHANGE DIRECTION
Why It Matters
A convincing setting or attractive pose cannot make a personalized image usable when the face is wrong, a foot is malformed, or an edit changes essential clothing.
In a sequence of images, success also depends on continuity. A satisfactory result needs to survive the next required change without reopening previously resolved problems.

What I Evaluated
I examined requested-edit completion, likeness preservation, anatomical plausibility, body proportions, clothing and scene continuity, communication accuracy, and the effort required to reach acceptance.
I also distinguished complete repair from partial repair, regression, concealment, tolerated deviation, and abandonment.
Study Design

Evaluation Framework
Each assessable dimension can be evaluated as failed, partially met, or met, with a separate not-assessable state.
The framework separates requested-edit success, preservation of essential constraints, whole-image usability, and the user’s acceptance decision. The report does not claim that every output received a numeric score.
Research Approach
Retrospective, naturalistic evaluation of an extended image-generation session using screenshot analysis, turn-level annotation, and correction-chain tracing.
The creative task developed during the interaction. This was an adaptive workflow rather than a controlled benchmark.

Task Design
Create a series of nautical photographs representing my likeness across changes in pose, clothing, framing, and setting.
Corrections addressed visible defects while attempting to preserve details that had already become satisfactory.

Evidence Corpus
A 133-page annotated nautical-session record containing 112 numbered turns.
A separate eight-page origin record documents the earlier attempts that prompted the investigation. It supplies context and is excluded from the nautical outcome totals.
Artifacts Produced
Full case-study report, structured annotation workbook, operational failure taxonomy, correction-chain analysis, and companion Lab Notes.
A TRACEABLE, MULTI-STAGE EVALUATION
Evidence Boundary
A numbered turn may contain multiple attempts or delivered images. Some screenshots show the next request rather than an isolated output.
Page references locate evidence in the retained record. They do not independently identify every generation attempt.
Evidence
Evidence 1 · A Completion Claim Outran the Result
Observed Behavior
At T22, the assistant stated that it had removed the lump below my left hand. My annotation recorded that the artifact remained visible.
My Analysis
The completion statement needed to be checked against the output rather than treated as evidence of success. Reporting an edit and visibly completing it are separate events.
REPAIR CLAIM UNVERIFIED
Evidence 3 · Concealment Enabled Acceptance
Observed Behavior
After repeated foot corrections, I asked for different shoes if the system could not produce a proper left foot. T69 delivered toe-covering flats and was accepted.
My Analysis
The shoe change was a useful workaround. Because the toes were covered, the result did not demonstrate successful repair of their anatomy.
ACCEPTED THROUGH CONCEALMENT
Evidence 2 · A Local Improvement Introduced Another Problem
Observed Behavior
At T10, scale was corrected while likeness drift and blur appeared. At T63, foot positioning became acceptable while body proportions changed.
My Analysis
These are repair-regression examples because the target improved while another requirement deteriorated. An unrelated defect that merely persists is incomplete recovery, not sufficient evidence of regression.
TARGET IMPROVED · PRESERVATION FAILED
EXAMPLES FROM THE RETAINED RECORD
Trace This Finding
An accepted image concealed the original defect rather than demonstrating that it had been repaired.
T54–T68 → SHOE-CHANGE REQUEST → T69 · Evidence summary
01 · ORIGINAL CORRECTION
At T54, the left-leg adjustment was finally acceptable, but I recorded two right feet. Subsequent requests targeted foot alignment, toe order, and the relationship between the foot and cushion.
Evidence references: T54–T57, source pages 65–68.
Evidence Boundary
The flats conceal the toes. Their underlying anatomy cannot be assessed from the accepted image.
This trace supports a finding about concealment and acceptance. It does not establish what the hidden foot structure would have looked like.
02 · CONTINUED FAILURE
The sequence included altered shoes instead of corrected alignment, a four-toe foot, missing feet, and further proportion problems. At T68, another request to reverse the toe configuration remained unsuccessful.
Evidence references: T55–T68, source pages 66–84.
03 · USER-REQUESTED WORKAROUND
I wrote:
“just change my shoes if you cant give me a proper left foot”
The assistant described replacing the sandals with red ballet flats, with the toes fully covered.
Evidence reference: T69, source page 85.
04 · MY ASSESSMENT
I accepted the result after the shoe change. This was acceptance through a workaround, not a demonstrated anatomical repair.
The successful production decision and the unresolved repair question must remain separately visible.
Evidence reference: T69 annotation and reproduced screenshot, report Figure 3.
SOURCE: Nautical-session annotated record, T54–T69, and full report Figures 2–3.
What Emerged
1
Acceptance Is Not a Repair Score
A user may accept an image because the original defect is hidden or because remaining deviations are tolerable.
2
Preservation Is Part of the Task
Correcting one feature is insufficient when the same edit damages likeness, anatomy, clothing, or another essential requirement.
2
3
User Effort Belongs in the Evaluation
Repeated reference uploads, reformulations, retries, and changes of direction reveal the work required to obtain a usable result.
3
KEY PATTERNS ACROSS THE RECORD
4
Transitions Need Interpretation
Moving to a new pose can follow success, compromise, or abandonment. A new request alone does not establish that the previous image passed.
4
Failure Taxonomy / Evaluation Framework
Failure Mode
Definition
Observed Behavior
Likeness Drift
Identifying features depart from the personal reference.
Repeated face corrections and reference resupply.
Anatomical Error
Incorrect limb, digit, joint, or body structure.
Extra legs, incorrect foot orientation, four toes, and calf-foot merging.
Repair Regression
The target improves while another essential requirement deteriorates.
Scale corrected while likeness drift appeared.
Overcorrection
An adjustment moves beyond the intended correction.
Head or feet became too small after size corrections.
Continuity Failure
Established clothing or scene details change unintentionally across edits.
Shoes, dress trim, or other accepted details changed.
Edit Noncompliance
The requested correction is absent, misapplied, or replaced by another change.
Shoes changed instead of foot alignment.
Completion Claim Mismatch
A completion statement is not supported by the visible result or evaluator assessment.
The assistant reported artifact removal while my note recorded it as still visible.
Opaque Rejection
A reported restriction lacks a supported explanation sufficient for troubleshooting.
Correction requests were followed by reported blocks without a verified cause.
Identity recurrence is coded as likeness drift. Workaround and abandonment are recovery outcomes, rather than interchangeable failure labels.
REQUEST → VERIFY TARGET → CHECK PRESERVATION → CLASSIFY RECOVERY → RECORD ACCEPTANCE
A REUSABLE GUIDE TO OBSERVED FAILURE MODES
Finding 1

Local Repair Did Not Ensure Whole-Image Acceptance
Evidence:
Repair regression is coded in six numbered turns. At T10 and T63, the requested target improved while another requirement deteriorated.
Implication:
Evaluate the target edit and preservation of essential details separately.
Finding 3

Identity Maintenance Required Recurring User Work
Evidence:
Likeness drift is coded in 45 turns. The notes explicitly count 42 personal-photo attachments across ten resupply events.
Implication:
Reference continuity and identity preservation should be measured across an editing sequence, alongside the burden placed on the user.
Finding 2

Accepted Results Included Compromise
Evidence:
Three of the 13 accepted turns involved compromise. T69 used concealment. T70 and T107 retained tolerated deviations.
Implication:
Report complete repair, accepted workaround, and accepted deviation as distinct outcomes.
Finding 4

Delivery Friction and Acceptance Coexisted
Evidence:
Thirteen turns contain a reported block or flag. T40, T99, and T100 also contain an eventual accepted result.
Implication:
A successful turn can still contain failed attempts. Turn outcomes and attempt-level friction require separate records.
FROM OBSERVATION TO IMPLICATION
Key Findings
Recommendations
1
Preserve an accessible record of accepted identity, clothing, proportions, and scene details.
Check the requested correction against the visible output before claiming completion.
1
2
Inspect neighboring anatomy, clothing boundaries, and support surfaces after local edits.
Compare each edited result with the last accepted state.
2
5
Avoid presenting an inferred moderation trigger as a verified diagnosis.
5
3
Separate complete repair from partial repair, regression, workaround, and abandonment.
Record tolerated deviations alongside acceptance.
3
6
Measure attempts, reference resupply, abandoned chains, and time to acceptance in follow-up testing.
6
4
Clarify poses when competing interpretations materially affect the result.
Distinguish reported restrictions from technical interruptions when that information is available.
4
TURN INSIGHT INTO ACTION
Product Team Handoff
Use This To…
Evaluate whether an image-editing workflow completes the requested correction while preserving previously satisfactory details.
Working Artifact
A 112-turn annotation workbook supported by source-page references, failure definitions, recovery distinctions, and the full case-study report.
Input → Output
Review the request, references, visible output, and subsequent user response. Record target-edit success, preservation failures, recovery outcome, acceptance, and user effort separately.
Evidence Boundary
One adaptive observational session. The workbook supports inspection of this record and development of follow-up tests. It does not establish a general image-model performance rate.
Limitations
This study examines one user’s adaptive session with one observed product workflow.
The underlying image model and internal routing were not independently established.
Likeness and usability judgments reflect my familiarity with my own appearance and creative requirements.
The record contains screenshots and retrospective annotations rather than complete tool logs. Rejected outputs are generally unavailable for inspection.
The 112 numbered turns are not 112 individually identified images or generation attempts. Outcome percentages do not establish model accuracy or typical customer success.
The 42 reference attachments are explicitly counted personal-photo attachments, not unique photographs or a complete count of every reference supplied.
Structured labels were drafted with AI assistance and reviewed by me. No second annotator or inter-rater agreement study was used.
The study does not establish why particular outputs were blocked or whether praise improved image stability.
IMPORTANT CONTEXT
Skills Demonstrated
Evidence Traceability
AI Evaluation
Correction-Chain Analysis
Recovery Evaluation
Failure Taxonomy Design
Visual Output Evaluation
Qualitative Research
Turn-Level Annotation
Technical Writing
Annotation Schema Design
Human–AI Interaction
UX Analysis
Researcher’s Note
I began by trying to make pictures.
As the session continued, I found myself checking whether each correction preserved the things that had finally become right. A pose could be lovely and the setting convincing, but if the face was wrong or the foot visibly malformed, I still could not use the photograph.
Sometimes I changed direction because I wanted another picture. Sometimes I changed direction because I had given up on the correction. Those transitions can look similar in a transcript, but they mean different things when evaluating recovery.
The shoe change made that distinction especially clear. I accepted the image because the new shoes covered the problem. That was a useful decision for finishing the picture, but it was not evidence that the system had repaired the foot.
The question became larger than whether the system could produce an attractive image. Could it make the requested correction, preserve what already worked, and accurately recognize what remained unresolved?
Suzanne Hansen




