Quality Drift in Generated Images

Issue: Generated Image Quality Degrades Over Multiple Generation Iterations or Long Sequences

Frequency: Common

Symptoms

• First generated images: high quality
• Later images: artifacts, blur, degradation
• Cumulative noise across iterations
• Model collapse in extended generation (>100 steps)

Root Cause Iterative generation (diffusion, autoregressive) accumulates noise across steps. Error correction is imperfect; each step’s small errors compound. Models don’t have explicit mechanisms to prevent quality drift over long sequences.

Example

Scenario: Generate 10 images of product variations
Image 1: Sharp, clear product
Image 10: Blurred, artifacts, malformed details
Impact: Later images unusable; high rejection rate

Key Statistics

• Quality degradation: 5-10% per generation step
• Mean quality at step 50: 60% of baseline
• Quality variance across sequence: high (>30%)

Mitigation Strategies

1. Noise Scheduling: Carefully tune diffusion schedule to prevent accumulation
2. Error Correction: Add refinement step between generations
3. State Reset: Periodically reset internal state to prevent drift
4. Anchor to Prompts: Re-condition on original prompt at regular intervals

Metrics

• Quality score decay over generations
• Artifact detection rate across sequence

Alerts

• Quality drop >20% vs. step 1 → P2

References

• Diffusion Model Stability
• Error Propagation in Generative Models