results/meeting-103-v2.1-clean/round5/5-4_thumb.png

<EFBFBD>PNG


IHDR<01>(<28><>X<03>IDATx<54><78><EFBFBD>1 <>Om
?<3F><00><>r<>=<3D>IEND<4E>B`<60>
-												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

											
										
										
											2026-05-20 00:03:37 +09:00
+								<EFBFBD>PNG
 IHDR<01>(<28><>X<03>IDATx<54><78><EFBFBD>1 <>Om
 								?<3F><00><>r<>=<3D>IEND<4E>B`<60>