A client asks for a continuation of last quarter’s hero asset. The artist who made it has moved on. The prompt is half-remembered, the model has updated twice, and the reference images are somewhere in a shared drive. The team rebuilds from scratch, the new version drifts, and the client asks why it looks different. This is the reproduction problem, and it is entirely a memory problem.
Why reproduction fails
- Prompts edited in place: the final prompt overwrote every intermediate that actually worked.
- Lost seeds and parameters: re-rendering without them is a new lottery ticket.
- Moved-on models: the same prompt on a new model version is a different asset.
- Missing references: style and character references detached from the output.
- Unknown approval state: the team reproduces a draft instead of the signed-off version.
The reproduction protocol
- Identify the approved version explicitly — status flag, not filename guess.
- Recover the full production record: references, prompt, model and tool versions, parameters, seed.
- Pin the original model and tool versions; do not substitute current ones silently.
- Re-render and compare side by side with the approved version at full resolution.
- Record the outcome: reproduced exactly, reproduced with drift (note causes), or unreproducible (note gaps).
Drift diagnosis
When the re-render looks wrong
| Symptom | Likely cause | Fix |
|---|---|---|
| Composition matches, texture differs | Model version moved on | Pin original model version |
| Same prompt, different layout | Seed lost or changed | Recover seed from production record |
| Style is close but off | Reference image missing | Attach canonical references to the identity |
| Everything matches except mood | Parameter drift (guidance, steps) | Copy full parameter set, not just the prompt |
Implementation checklist
Reproducibility checklist
- Store seeds and full parameter sets with every review candidate.
- Pin model and tool versions in the production record.
- Never edit prompts in place — each revision is a new version with its own prompt.
- Verify reproductions side by side before calling them done.
- Promote reproduced assets to reusable recipes for the next brief.
What to measure in a paid pilot
Track reproduction success without the original artist, mean time to reproduce, and drift incidents per campaign during a production pilot. The Reproduce AI-generated assets workflow shows what to measure.
Start with the workflow assessment to find where your team loses context today, read the Production Memory guide, and browse the use-case library for related workflows.
Key takeaways
- Reproduction = recover the production record, re-render, verify against approved.
- The five usual gaps: edited-in-place prompts, lost seeds, moved-on models, missing references, unknown approval state.
- Verify with side-by-side review, not memory — near-identical is not identical.
- Reproducible assets become reusable recipes for future campaigns.