Skip to content

Case 03: Avedon Project

All case studies are observational logs from real production workflows.\ Results may vary by model version and configuration.

Scenario: Skeletal-level identity control for fashion industry standards
Platform: ChatGPT (GPT Image 1.5)
Protocol: Advanced anchor + formal validation gates
Result: Audit-ready consistency observed across 38 generation cycles (77 total exchanges) — identity preserved through lighting, material, and monochrome transitions

Turn counting: 77 total exchanges (39 user instructions + 38 AI generations). "38 turns" = 38 generation cycles.

What This Demonstrates

This is the most technically demanding case study.

A ballerina character (Model "A") is held at a fixed pose across 38 generation turns.\ The operator changes only one condition per turn: lighting angle, material density, exposure, framing, and monochrome conversion.\ Identity, skeletal alignment, and pose are required to remain constant throughout.

This case suggests the protocol can meet fashion/editorial production requirements --- where repeatability is often contractually required.

Turn Accounting Clarification

Terms used in this document:

Term Definition
Generation turn One user instruction + one AI image generation (a pair)
Evaluation step Human match rate assessment performed after each generation turn
Stabilization checkpoint Re-anchoring event — anchor image re-injected to recover convergence

Numeric summary for this session: - Total exchanges: 77 (39 user instructions + 38 AI generations) - Generation turns: 38 - Evaluation steps: 38 (one per generation turn) - Stabilization checkpoints: 1 (Turn 29)

These definitions apply consistently across this document, the whitepaper, and all README references.

Runs were performed under single-session conditions.\ Cross-session behavior may vary by model version and system state.

Session Structure

Phase Turns What Changed

Anchor 1--10 Identity lock, pose anchor set establishment

Skeletal control 11--30 Pose definition tightened instruction by instruction

Lighting variation 31--54 Side light, backlight, exposure, framing

Material variation 55--66 Skirt material, sheen, specular

Monochrome + 67--77 Grayscale conversion, background background variation

Analysis

Metric Result

Total turns 38 generation cycles (77 total exchanges)

Pose changes 0 (frozen throughout)

Conditions varied Lighting × 6, Material × 3, Monochrome × 2, Background × 1

Match rate threshold 90% (session abandoned if breached)

Re-anchoring events 1 (Turn 29)

Session outcome Prompt fixed, contaminated generations purged

Key finding: Identity consistency was preserved across tested condition variations through skeletal-level prompt control. When deviation was detected (Turn 77), the session was formally terminated rather than continued with degraded output.

Source

  • Case 01: Baseline Failure --- same workflow without protocol\
  • Technical Mechanism --- how anchor control works