Artifact Accumulation in Regeneration

Issue: Repeated Generation/Regeneration of Same Content Introduces Synthetic Artifacts

Frequency: Common

Symptoms

• Regenerated images contain unnatural patterns, distortions
• Artifacts cluster in specific regions (eyes, hands, textures)
• Quality decreases with each regeneration round
• Visible “compression” or “bleaching” artifacts

Root Cause Generation models learn from training data that rarely contains regenerated/resampled images. Repeated generation exposes model to distribution mismatch — the model’s own outputs are not well-represented in training data. Leads to increasingly unnatural patterns.

Example

Scenario: Iterative product image refinement
Initial: Photo-realistic product image
Regenerate to "rotate 10 degrees": Artifacts appear around edges
Regenerate again: More artifacts, distortion amplifies
Impact: Eventually unusable image

Key Statistics

• Artifact detection: 10% (1st gen) → 40% (3rd gen)
• User rejection rate: 5% (1st) → 30% (3rd)

Mitigation Strategies

1. One-Shot Generation: Prefer single generation over iterative refinement
2. Regeneration Threshold: Limit regeneration rounds (max 2-3)
3. Artifact Detection: Run artifact detector before accepting regenerated image
4. Human Review: Manual approval for regenerated content

Metrics

• Artifact detection rate per generation round
• User rejection rate per round

Alerts

• Artifact rate >25% → Reject regeneration

References

• Synthetic Artifact Detection in Generative Models
• Mode Coverage in Diffusion Models