meeting-103 v2.1 clean R5+R6 ✅: byeolyi-default.png ref = pastel 69.66% (+13pp 점프)

R5 (3/5 success):
- 5-1: Stage 3 Refiner denoise 0.25 → 49.85% (효과 미미)
- 5-2: PuLID 0.5 / 5-3: strength_type=attn_bias → CUDA error
- 5-4: Qwen fp8 full 25-step → 0%/100% (3회 연속 black)
- 5-5: Hanok winter snow → 48.65% (spring/winter > summer)

R6 (4/5 success, 별이 huge best 발견):
- 6-1: ref byeolyi-916.png → 52.99% (-3.4pp vs 6-2)
- 6-2: ref byeolyi-default.png → ★★★ 69.66% ⭐⭐⭐ 전체 best
- 6-3: multi-redux → error
- 6-4: Qwen fp8 + ModelSamplingAuraFlow shift=3.0 → 0%/100% (4회째 black)
- 6-5: Hanok dancheong (Korean color cues) → 15.09%

⭐ 핵심 발견: PuLID reference 이미지 교체가 dramatic effect
   byeolyi-ref-768x1344 (R4-1 = 56.41%) → byeolyi-default (R6-2 = 69.66%) = +13.25pp

진행: R9 legacy 44.5% → R1 48.5% → R2 49.99% → R3 50.47% → R4 56.41% → R6 69.66%

Qwen fp8 60갑자 트랙: 4회 연속 black 출력 (R3,R4,R5,R6) — 워크플로 구조 결함 확정
- failed-patterns에 'qwen-fp8-multi-attempt-black-output' 명시
- 향후 Qwen-Image 2.0 native (non-distill) 시도 필요

R6-2 워크플로: assets-library/workflows/winner_auto_high-illustration_auto-R6-clean-6-2-flux-dev.json

assets-library/prompts/failed-patterns.json

@@ -114,5 +114,21 @@
       "variant": "Qwen fp8 + Lightning LoRA 4-step lcm/normal cfg=1.0",
       "note": "pastel=0% black=100% — lcm scheduler 도 해결 X, Lightning LoRA strength_clip=1.0 부적절 의심"
     }
+  ],
+  "qwen-fp8-FULL-25step-eulerA-cfg25-black-output-R5": [
+    {
+      "round": "R5-clean",
+      "id": "5-4",
+      "variant": "Qwen fp8 (no Lightning) 25-step euler/normal cfg=2.5",
+      "note": "pastel=0% black=100% — Qwen-Image fp8 워크플로 구조 결함 (ModelSamplingAuraFlow shift 노드 누락 의심). 3회 연속 black 출력 (R3-4, R4-4, R5-4 동일)"
+    }
+  ],
+  "qwen-fp8-ModelSamplingAuraFlow-shift3-still-black-R6": [
+    {
+      "round": "R6-clean",
+      "id": "6-4",
+      "variant": "Qwen fp8 + ModelSamplingAuraFlow shift=3.0 + 30 step euler/simple cfg=2.5",
+      "note": "pastel=0% black=100% — ModelSamplingAuraFlow도 해결 X. 4회 연속 black (R3,R4,R5,R6). Qwen fp8 워크플로 구조 근본 결함 - 다른 접근 필요 (Qwen-Image 2.0 native sampler? text_encoders 분리?)"
+    }
   ]
 }