Consistency Failure: Identity Loss Across Generations

Issue: Generated Images Lack Consistency Across Multiple Generations (Same Subject, Different Outputs)

Frequency: Common

Symptoms

• Same prompt generates visually different outputs (good for diversity)
• But can be bad when consistency is needed (character, product)
• Object identity shifts between frames
• Character appearance drastically changes across images

Root Cause Generative models sample stochastically; they have no built-in mechanism to maintain identity/consistency across samples. This is a feature for diversity but a bug when consistency is needed (animation, character design, product continuity).

Example

Scenario: Animated character generation
Prompt: "Generate 10 frames of character walking"

Frame 1: Character with brown hair, blue shirt
Frame 2: Character with red hair, yellow shirt
Frame 3: Character with blonde hair, green shirt
Impact: Incoherent animation; perceived as different characters

Key Statistics

• Identity preservation: 40-60% across 10 generations
• Attribute drift: ~5% per frame on average

Mitigation Strategies

1. Seed Reuse: Use same random seed across variations (for deterministic parts)
2. Identity Embedding: Extract identity token; reuse across generations
3. LoRA Adaptation: Fine-tune identity-specific model variant
4. Reference Image: Condition each generation on previous frame

Metrics

• LPIPS distance (perceptual similarity between frames)
• Attribute consistency (face recognition: same identity across images)

Alerts

• Identity score <50% → P2

References

• Consistent Image Synthesis
• Identity-Preserving Generative Models