Built-in imagegen produced unrelated outputs across retries despite tight negative constraints

Summary

The built-in image_gen tool in Codex produced multiple unrelated outputs for the same creative image task, even after tightening the prompt with explicit negative constraints. A separate chat/session was able to produce an on-target result for the same concept family, so this looks like intermittent imagegen routing, context carryover, prompt adherence, or retry-state drift rather than a bad concept prompt alone.

In-app feedback was already uploaded.

Feedback ID:

019eed6e-82f4-7043-8bf5-bf75981f2b6a

Related earlier feedback ID from a different chat/session:

019ee5e3-6fb4-76f2-9a7b-ff6da7ae2372

Possibly related issue, but not the same failure mode:

• #22116 covers hatch-pet workflow drift plus stream disconnection.
• This report covers imagegen output drifting to unrelated image categories across retries.

Expected Behavior

When the user asks for a fictional "Little Buddy / context farmer" creative image, the generated image should remain in that semantic neighborhood:

• fictional helper/buddy subject
• desk-garden or context-farm scene
• abstract device/photo/system layers
• no readable private UI
• no unrelated politics/news/history content
• no unrelated photoreal restaurant/person output
• no generic self-help poster unless requested

Actual Behavior

Three consecutive attempts drifted into unrelated content:

1. A generic "10 daily habits" wellness poster.
2. A photoreal restaurant/ramen portrait scene.
3. A Middle East conflict timeline infographic.

The third result is especially concerning because the prompt explicitly prohibited politics, war, news, flags, maps, and timeline/infographic formats.

Additional evidence from the earlier chat/session:

• A reporting / spreadsheet concept image task drifted into unrelated self-care and chemistry education posters.
• A later attempt in that same general workflow did produce an on-target reporting-focused concept image, which suggests the issue may be intermittent rather than purely prompt quality.
• Another visible drift example in the evidence set showed unrelated civic/activities educational posters despite a reporting-app target.

Sanitized Prompt Shape

The prompt was a creative-local image request. It asked for:

• wide 16:9 single cinematic desk diorama
• fictional toy-like explorer buddy in a turquoise-blue astronaut suit
• context farm / desk-garden metaphor
• abstract computer/phone/photo panes
• small everyday symbolic props
• Taylor-series curves, three-body orbits, rings, vectors, and a purple-green moon
• hard negatives against infographic, poster, timeline, readable UI, logos, politics, news imagery, real screenshots, and real private photos

The final tightened prompt began with:

Single cinematic fantasy diorama on a wooden desk. No infographic. No poster. No panels. No numbered list. No timeline. No headline. No readable text anywhere. No flags, maps, politics, war, news, or corporate logos.

Why This Seems Like A Bug

• Tightening the prompt did not reduce drift.
• The output categories were not adjacent misses; they were different task families.
• One failure directly violated hard negative constraints.
• The same broad concept did work in a different chat/session, suggesting non-deterministic tool/session behavior rather than only prompt ambiguity.
• A separate reporting-image task showed similar category drift, then later produced a successful on-target result.
• The app accepted in-product feedback and returned Feedback IDs, so there are traceable internal artifacts.

Privacy / Safety Notes

This issue intentionally omits raw screenshots, private photo libraries, local file paths, and any private prompt context. The included details are sanitized summaries only. The Feedback IDs above can be used to correlate with uploaded in-app feedback if maintainers have access.

Acceptance Criteria

• Imagegen retries should stay within the requested subject/domain when prompts include strong subject anchors and hard negatives.
• Hard negative constraints against politics/news/timelines should not produce a conflict-history timeline.
• If imagegen routing/context becomes stale or contaminated, Codex should reset the generation context or expose a retry/reset control.
• If the image tool returns an unrelated task family, Codex should surface enough metadata or error state to distinguish model drift from a successful generation.