choijaewook/8460s-image-rd
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
74a0415e8cd5b305066aa6fd80ba48a1577159b6
8460s-image-rd/eval/meeting-103-v2.1-clean-grid.html

198 lines
13 KiB
HTML
Raw Normal View History

meeting-103 v2.1 clean R1 ✅5/5: PuLID + Redux = pastel 48.5% (R9 best 갱신) R0 환경 점검: - VRAM 11.7/12.4GB free (94%) — 해소 확인 - pytorch 2.12.0+cu130 + sageattention venv 검증 - 모델 인벤토리: Flux Dev Q5 GGUF, PuLID, Redux, byeolyi-ref-916 확보 - fp8 다운로드 .part 0 bytes (재개 대기) R1 clean 5조합 (5/5 success): - 1-1: PuLID w0.85 mean → pastel 25.7% / black 1.95% - 1-2: PuLID w0.85 mean delayed_start 0.05 → pastel 10.0% / black 14.0% (어두움) - 1-3: PuLID w1.0 mean → pastel 27.5% / black 0.79% - 1-4: PuLID w0.7 + Redux 0.5 → pastel 48.5% / black 0.27% ★ R1 best - 1-5: 한옥 baseline Flux Dev 768² → pastel 12.9% / black 11.7% CUDA error 회복: - 1-2 retry concat → CUDA invalid argument → 후속 1-2~1-5 모두 실패 - ComfyUI venv 재시작 (sage + pytorch 2.12) → 1-2 변형 (delayed_start)로 대체 - 재시도 4/4 모두 success 이전 R9 best (44.5%) → R1-4 (48.5%) = +4 pp 갱신 자산 라이브러리 누적: - models-performance.json: 7 모델 추적 (flux-dev, pulid-flux, flux-redux 등) - workflows: winner_auto 1 + partial 2 - prompts/proven: illustration_auto 패턴 3건 - prompts/failed: pastel-low + CUDA error 패턴 - prompts/negative: A_byeolyi 패턴 etc. 검수: - Vault: D:\Vault\8460s-image-rd\eval\meeting-103-v2.1-clean-grid.html - Gitea: results/meeting-103-v2.1-clean/round1/
2026-05-19 23:17:19 +09:00
<!DOCTYPE html>
<html lang="ko"><head><meta charset="utf-8">
meeting-103 v2.1 clean R3+R4 ✅: PuLID 0.7+Redux 1.0 = pastel 56.41% (전체 best 갱신) R3 (4/5 success): - 3-1: PuLID 0.7 + Redux 0.85 → 50.47% (R2 best 49.99% 갱신) - 3-2: PuLID 0.6 + Redux 0.7 → 49.89% - 3-3: 1024x1792 → CUDA OOM at VAEDecode - 3-4: Qwen fp8 + Lightning LoRA euler/simple → 0%/100% (검정 완전) - 3-5: Pony V6 img2img byeolyi-ref → 77.3% (img2img 0.65 denoise, 원본 색 보존) R4 (3/5 success, 별이 새 best): - 4-1: PuLID 0.7 + Redux 1.0 (max) → 56.41% ★ 전체 best - 4-2: PuLID 0.85 + Redux 0.7 → 50.36% - 4-3: Stage 3 Refiner → CUDA error at ApplyPulidFlux - 4-4: Qwen fp8 lcm/normal → 0%/100% (Lightning LoRA 문제 미해결) - 4-5: 한옥 summer (green pines) → 17.7%/3.97% 학습: - Redux strength 0.5 → 0.7 → 0.85 → 1.0 단조 증가 (56.41% sweet spot) - PuLID 0.7 fixed, Redux 1.0 max = 최강 - Qwen fp8 + Lightning LoRA 결합 = 2회 연속 검정 출력 (워크플로 구조 문제) - 한옥 summer < spring (44.45%) — 색감 자체가 차분 진행: R9 legacy 44.5% → R1 48.5% → R2 49.99% → R3 50.47% → R4 56.41% (+11.91pp) 자산 라이브러리: - workflows: winner_auto 7 + partial 7 - prompts/failed: 9 패턴 (Qwen 계열 2건 추가)
2026-05-19 23:49:55 +09:00
<title>meeting-103 v2.1 clean — R1~R4</title>
meeting-103 v2.1 clean R1 ✅5/5: PuLID + Redux = pastel 48.5% (R9 best 갱신) R0 환경 점검: - VRAM 11.7/12.4GB free (94%) — 해소 확인 - pytorch 2.12.0+cu130 + sageattention venv 검증 - 모델 인벤토리: Flux Dev Q5 GGUF, PuLID, Redux, byeolyi-ref-916 확보 - fp8 다운로드 .part 0 bytes (재개 대기) R1 clean 5조합 (5/5 success): - 1-1: PuLID w0.85 mean → pastel 25.7% / black 1.95% - 1-2: PuLID w0.85 mean delayed_start 0.05 → pastel 10.0% / black 14.0% (어두움) - 1-3: PuLID w1.0 mean → pastel 27.5% / black 0.79% - 1-4: PuLID w0.7 + Redux 0.5 → pastel 48.5% / black 0.27% ★ R1 best - 1-5: 한옥 baseline Flux Dev 768² → pastel 12.9% / black 11.7% CUDA error 회복: - 1-2 retry concat → CUDA invalid argument → 후속 1-2~1-5 모두 실패 - ComfyUI venv 재시작 (sage + pytorch 2.12) → 1-2 변형 (delayed_start)로 대체 - 재시도 4/4 모두 success 이전 R9 best (44.5%) → R1-4 (48.5%) = +4 pp 갱신 자산 라이브러리 누적: - models-performance.json: 7 모델 추적 (flux-dev, pulid-flux, flux-redux 등) - workflows: winner_auto 1 + partial 2 - prompts/proven: illustration_auto 패턴 3건 - prompts/failed: pastel-low + CUDA error 패턴 - prompts/negative: A_byeolyi 패턴 etc. 검수: - Vault: D:\Vault\8460s-image-rd\eval\meeting-103-v2.1-clean-grid.html - Gitea: results/meeting-103-v2.1-clean/round1/
2026-05-19 23:17:19 +09:00
<style>
meeting-103 v2.1 clean R3+R4 ✅: PuLID 0.7+Redux 1.0 = pastel 56.41% (전체 best 갱신) R3 (4/5 success): - 3-1: PuLID 0.7 + Redux 0.85 → 50.47% (R2 best 49.99% 갱신) - 3-2: PuLID 0.6 + Redux 0.7 → 49.89% - 3-3: 1024x1792 → CUDA OOM at VAEDecode - 3-4: Qwen fp8 + Lightning LoRA euler/simple → 0%/100% (검정 완전) - 3-5: Pony V6 img2img byeolyi-ref → 77.3% (img2img 0.65 denoise, 원본 색 보존) R4 (3/5 success, 별이 새 best): - 4-1: PuLID 0.7 + Redux 1.0 (max) → 56.41% ★ 전체 best - 4-2: PuLID 0.85 + Redux 0.7 → 50.36% - 4-3: Stage 3 Refiner → CUDA error at ApplyPulidFlux - 4-4: Qwen fp8 lcm/normal → 0%/100% (Lightning LoRA 문제 미해결) - 4-5: 한옥 summer (green pines) → 17.7%/3.97% 학습: - Redux strength 0.5 → 0.7 → 0.85 → 1.0 단조 증가 (56.41% sweet spot) - PuLID 0.7 fixed, Redux 1.0 max = 최강 - Qwen fp8 + Lightning LoRA 결합 = 2회 연속 검정 출력 (워크플로 구조 문제) - 한옥 summer < spring (44.45%) — 색감 자체가 차분 진행: R9 legacy 44.5% → R1 48.5% → R2 49.99% → R3 50.47% → R4 56.41% (+11.91pp) 자산 라이브러리: - workflows: winner_auto 7 + partial 7 - prompts/failed: 9 패턴 (Qwen 계열 2건 추가)
2026-05-19 23:49:55 +09:00
*{box-sizing:border-box}body{font-family:-apple-system,'Apple SD Gothic Neo',sans-serif;background:#0e0e15;color:#eaeaf0;padding:24px}h1{margin:0 0 8px;font-size:22px}.subtitle{color:#9b9bb0;margin-bottom:16px;font-size:13px}.summary{background:#1c1c28;border-left:3px solid #6f7df2;padding:12px 16px;margin-bottom:24px;border-radius:6px;line-height:1.6}.summary b{color:#c8caff}.grid{display:grid;grid-template-columns:repeat(auto-fill,minmax(360px,1fr));gap:16px}.cell{background:#18181f;border:1px solid #2a2a36;border-radius:8px;overflow:hidden}.cell.status-error{border-color:#5a2222;background:#1c1216}.cell.status-success{border-color:#2a4a2a}.cell img{width:100%;display:block;max-height:540px;object-fit:contain;background:#0a0a10}.no-image{width:100%;padding:60px 12px;text-align:center;color:#c47;background:#1c1216;font-weight:bold}.label{padding:10px 12px;font-size:12px}.label-title{font-weight:600;color:#d4d4e0;margin-bottom:4px}.label-meta{color:#8a8aa0;font-family:ui-monospace,monospace;margin:4px 0}.label-auto-eval{color:#b0c4b0;margin:6px 0;line-height:1.45}.label-flags{margin-top:6px;display:flex;flex-wrap:wrap;gap:4px}.flag{display:inline-block;padding:2px 6px;border-radius:3px;font-size:10px;font-weight:600}.flag-ok{background:#1f4220;color:#8fcd8f}.flag-warn{background:#4a2a16;color:#f0a96a}
meeting-103 v2.1 clean R1 ✅5/5: PuLID + Redux = pastel 48.5% (R9 best 갱신) R0 환경 점검: - VRAM 11.7/12.4GB free (94%) — 해소 확인 - pytorch 2.12.0+cu130 + sageattention venv 검증 - 모델 인벤토리: Flux Dev Q5 GGUF, PuLID, Redux, byeolyi-ref-916 확보 - fp8 다운로드 .part 0 bytes (재개 대기) R1 clean 5조합 (5/5 success): - 1-1: PuLID w0.85 mean → pastel 25.7% / black 1.95% - 1-2: PuLID w0.85 mean delayed_start 0.05 → pastel 10.0% / black 14.0% (어두움) - 1-3: PuLID w1.0 mean → pastel 27.5% / black 0.79% - 1-4: PuLID w0.7 + Redux 0.5 → pastel 48.5% / black 0.27% ★ R1 best - 1-5: 한옥 baseline Flux Dev 768² → pastel 12.9% / black 11.7% CUDA error 회복: - 1-2 retry concat → CUDA invalid argument → 후속 1-2~1-5 모두 실패 - ComfyUI venv 재시작 (sage + pytorch 2.12) → 1-2 변형 (delayed_start)로 대체 - 재시도 4/4 모두 success 이전 R9 best (44.5%) → R1-4 (48.5%) = +4 pp 갱신 자산 라이브러리 누적: - models-performance.json: 7 모델 추적 (flux-dev, pulid-flux, flux-redux 등) - workflows: winner_auto 1 + partial 2 - prompts/proven: illustration_auto 패턴 3건 - prompts/failed: pastel-low + CUDA error 패턴 - prompts/negative: A_byeolyi 패턴 etc. 검수: - Vault: D:\Vault\8460s-image-rd\eval\meeting-103-v2.1-clean-grid.html - Gitea: results/meeting-103-v2.1-clean/round1/
2026-05-19 23:17:19 +09:00
</style></head><body>
meeting-103 v2.1 clean R3+R4 ✅: PuLID 0.7+Redux 1.0 = pastel 56.41% (전체 best 갱신) R3 (4/5 success): - 3-1: PuLID 0.7 + Redux 0.85 → 50.47% (R2 best 49.99% 갱신) - 3-2: PuLID 0.6 + Redux 0.7 → 49.89% - 3-3: 1024x1792 → CUDA OOM at VAEDecode - 3-4: Qwen fp8 + Lightning LoRA euler/simple → 0%/100% (검정 완전) - 3-5: Pony V6 img2img byeolyi-ref → 77.3% (img2img 0.65 denoise, 원본 색 보존) R4 (3/5 success, 별이 새 best): - 4-1: PuLID 0.7 + Redux 1.0 (max) → 56.41% ★ 전체 best - 4-2: PuLID 0.85 + Redux 0.7 → 50.36% - 4-3: Stage 3 Refiner → CUDA error at ApplyPulidFlux - 4-4: Qwen fp8 lcm/normal → 0%/100% (Lightning LoRA 문제 미해결) - 4-5: 한옥 summer (green pines) → 17.7%/3.97% 학습: - Redux strength 0.5 → 0.7 → 0.85 → 1.0 단조 증가 (56.41% sweet spot) - PuLID 0.7 fixed, Redux 1.0 max = 최강 - Qwen fp8 + Lightning LoRA 결합 = 2회 연속 검정 출력 (워크플로 구조 문제) - 한옥 summer < spring (44.45%) — 색감 자체가 차분 진행: R9 legacy 44.5% → R1 48.5% → R2 49.99% → R3 50.47% → R4 56.41% (+11.91pp) 자산 라이브러리: - workflows: winner_auto 7 + partial 7 - prompts/failed: 9 패턴 (Qwen 계열 2건 추가)
2026-05-19 23:49:55 +09:00
<h1>meeting-103 v2.1 clean — R1~R4 그리드</h1>
<p class="subtitle">총 20 조합 / 성공 18 · 진행 라운드 4</p>
meeting-103 v2.1 clean R1 ✅5/5: PuLID + Redux = pastel 48.5% (R9 best 갱신) R0 환경 점검: - VRAM 11.7/12.4GB free (94%) — 해소 확인 - pytorch 2.12.0+cu130 + sageattention venv 검증 - 모델 인벤토리: Flux Dev Q5 GGUF, PuLID, Redux, byeolyi-ref-916 확보 - fp8 다운로드 .part 0 bytes (재개 대기) R1 clean 5조합 (5/5 success): - 1-1: PuLID w0.85 mean → pastel 25.7% / black 1.95% - 1-2: PuLID w0.85 mean delayed_start 0.05 → pastel 10.0% / black 14.0% (어두움) - 1-3: PuLID w1.0 mean → pastel 27.5% / black 0.79% - 1-4: PuLID w0.7 + Redux 0.5 → pastel 48.5% / black 0.27% ★ R1 best - 1-5: 한옥 baseline Flux Dev 768² → pastel 12.9% / black 11.7% CUDA error 회복: - 1-2 retry concat → CUDA invalid argument → 후속 1-2~1-5 모두 실패 - ComfyUI venv 재시작 (sage + pytorch 2.12) → 1-2 변형 (delayed_start)로 대체 - 재시도 4/4 모두 success 이전 R9 best (44.5%) → R1-4 (48.5%) = +4 pp 갱신 자산 라이브러리 누적: - models-performance.json: 7 모델 추적 (flux-dev, pulid-flux, flux-redux 등) - workflows: winner_auto 1 + partial 2 - prompts/proven: illustration_auto 패턴 3건 - prompts/failed: pastel-low + CUDA error 패턴 - prompts/negative: A_byeolyi 패턴 etc. 검수: - Vault: D:\Vault\8460s-image-rd\eval\meeting-103-v2.1-clean-grid.html - Gitea: results/meeting-103-v2.1-clean/round1/
2026-05-19 23:17:19 +09:00
<div class="summary">
meeting-103 v2.1 clean R3+R4 ✅: PuLID 0.7+Redux 1.0 = pastel 56.41% (전체 best 갱신) R3 (4/5 success): - 3-1: PuLID 0.7 + Redux 0.85 → 50.47% (R2 best 49.99% 갱신) - 3-2: PuLID 0.6 + Redux 0.7 → 49.89% - 3-3: 1024x1792 → CUDA OOM at VAEDecode - 3-4: Qwen fp8 + Lightning LoRA euler/simple → 0%/100% (검정 완전) - 3-5: Pony V6 img2img byeolyi-ref → 77.3% (img2img 0.65 denoise, 원본 색 보존) R4 (3/5 success, 별이 새 best): - 4-1: PuLID 0.7 + Redux 1.0 (max) → 56.41% ★ 전체 best - 4-2: PuLID 0.85 + Redux 0.7 → 50.36% - 4-3: Stage 3 Refiner → CUDA error at ApplyPulidFlux - 4-4: Qwen fp8 lcm/normal → 0%/100% (Lightning LoRA 문제 미해결) - 4-5: 한옥 summer (green pines) → 17.7%/3.97% 학습: - Redux strength 0.5 → 0.7 → 0.85 → 1.0 단조 증가 (56.41% sweet spot) - PuLID 0.7 fixed, Redux 1.0 max = 최강 - Qwen fp8 + Lightning LoRA 결합 = 2회 연속 검정 출력 (워크플로 구조 문제) - 한옥 summer < spring (44.45%) — 색감 자체가 차분 진행: R9 legacy 44.5% → R1 48.5% → R2 49.99% → R3 50.47% → R4 56.41% (+11.91pp) 자산 라이브러리: - workflows: winner_auto 7 + partial 7 - prompts/failed: 9 패턴 (Qwen 계열 2건 추가)
2026-05-19 23:49:55 +09:00
전체 best pastel: <b>R3-3-5</b> Pony V6 img2img from byeolyi-ref (denoise 0.65) = 77.3%<br>
라운드별 best: R1: 48.50% (1-4) / R2: 49.99% (2-4) / R3: 77.30% (3-5) / R4: 56.41% (4-1)<br>
R9 legacy 44.5% → R1 48.5% → R2 49.99% → R3 50.47% → <b>R4 56.41% (Redux 1.0 sweet spot)</b><br>
⚠ vision 검수 대기 (별이≥8 / 한자≥4.5 / 한옥≥4.5) — auto 지표는 proxy
meeting-103 v2.1 clean R1 ✅5/5: PuLID + Redux = pastel 48.5% (R9 best 갱신) R0 환경 점검: - VRAM 11.7/12.4GB free (94%) — 해소 확인 - pytorch 2.12.0+cu130 + sageattention venv 검증 - 모델 인벤토리: Flux Dev Q5 GGUF, PuLID, Redux, byeolyi-ref-916 확보 - fp8 다운로드 .part 0 bytes (재개 대기) R1 clean 5조합 (5/5 success): - 1-1: PuLID w0.85 mean → pastel 25.7% / black 1.95% - 1-2: PuLID w0.85 mean delayed_start 0.05 → pastel 10.0% / black 14.0% (어두움) - 1-3: PuLID w1.0 mean → pastel 27.5% / black 0.79% - 1-4: PuLID w0.7 + Redux 0.5 → pastel 48.5% / black 0.27% ★ R1 best - 1-5: 한옥 baseline Flux Dev 768² → pastel 12.9% / black 11.7% CUDA error 회복: - 1-2 retry concat → CUDA invalid argument → 후속 1-2~1-5 모두 실패 - ComfyUI venv 재시작 (sage + pytorch 2.12) → 1-2 변형 (delayed_start)로 대체 - 재시도 4/4 모두 success 이전 R9 best (44.5%) → R1-4 (48.5%) = +4 pp 갱신 자산 라이브러리 누적: - models-performance.json: 7 모델 추적 (flux-dev, pulid-flux, flux-redux 등) - workflows: winner_auto 1 + partial 2 - prompts/proven: illustration_auto 패턴 3건 - prompts/failed: pastel-low + CUDA error 패턴 - prompts/negative: A_byeolyi 패턴 etc. 검수: - Vault: D:\Vault\8460s-image-rd\eval\meeting-103-v2.1-clean-grid.html - Gitea: results/meeting-103-v2.1-clean/round1/
2026-05-19 23:17:19 +09:00
</div>
<div class="grid">
<div class="cell status-success">
<img src="../results/meeting-103-v2.1-clean/round1/1-1_thumb.png" data-original="../results/meeting-103-v2.1-clean/round1/1-1.png" alt="1-1">
<div class="label">
meeting-103 v2.1 clean R2 ✅5/5: PuLID 0.85+Redux 0.5 = pastel 49.99% (전체 best 갱신) R2 5조합 (5/5 success, retry 2회 포함): - 2-1: PuLID 0.7 + Redux 0.5 + Stage 3 Refiner (denoise 0.3) → 48.82% (Stage 3 효과 미미) - 2-2: PuLID 0.7 + Redux 0.7 → 48.93% - 2-3: PuLID 0.7 + Redux 0.3 → 38.52% (Redux 낮으면 파스텔 감소) - 2-4: PuLID 0.85 + Redux 0.5 → 49.99% ★ R2 best & 전체 best - 2-5: 한옥 spring (cherry blossom) → 44.45% / black 4.3% R1 best (R1-4 48.5%) → R2-4 (49.99%) = +1.5 pp 갱신 R9 legacy (44.5%) → R2-4 (49.99%) = +5.5 pp 누적 학습: - PuLID weight 0.85 + Redux 0.5 multiply 조합이 0.7/0.5 보다 우월 - Redux 0.3 너무 낮음, 0.7 비슷, 0.85+ 미검증 - Stage 3 Refiner 효과 미미 (48.5% → 48.82%, +0.3pp) CUDA error 회복: - 2-3, 2-4, 2-5 첫 시도 CUDA invalid argument - ComfyUI venv 재시작 후 retry → 3/3 모두 success - 매 ~5건 후 재시작 패턴 확인됨 자산 라이브러리: - models-performance: 8 모델 추적 - workflows: winner_auto 4 (R1-4, R2-1, R2-2, R2-4) + partial 5 - prompts/proven: 7 패턴 (R1, R2 합산) R3 계획: - 3-1: PuLID 0.7 + Redux 0.85 (push redux 더 강) - 3-2: PuLID 0.6 + Redux 0.7 (PuLID 약화) - 3-3: PuLID 0.7 + Redux 0.7 @ 1024x1792 (해상도 증가) - 3-4: Qwen-Image fp8 + Lightning LoRA 60갑자 (fp8 ~20G 완료, GGUF 회피) - 3-5: Pony V6 img2img from byeolyi-ref (대안 ID 방법) 검수: D:\Vault\8460s-image-rd\eval\meeting-103-v2.1-clean-grid.html
2026-05-19 23:36:44 +09:00
<div class="label-title">R1-1-1: A — PuLID weight 0.85 mean (first try)</div>
meeting-103 v2.1 clean R3+R4 ✅: PuLID 0.7+Redux 1.0 = pastel 56.41% (전체 best 갱신) R3 (4/5 success): - 3-1: PuLID 0.7 + Redux 0.85 → 50.47% (R2 best 49.99% 갱신) - 3-2: PuLID 0.6 + Redux 0.7 → 49.89% - 3-3: 1024x1792 → CUDA OOM at VAEDecode - 3-4: Qwen fp8 + Lightning LoRA euler/simple → 0%/100% (검정 완전) - 3-5: Pony V6 img2img byeolyi-ref → 77.3% (img2img 0.65 denoise, 원본 색 보존) R4 (3/5 success, 별이 새 best): - 4-1: PuLID 0.7 + Redux 1.0 (max) → 56.41% ★ 전체 best - 4-2: PuLID 0.85 + Redux 0.7 → 50.36% - 4-3: Stage 3 Refiner → CUDA error at ApplyPulidFlux - 4-4: Qwen fp8 lcm/normal → 0%/100% (Lightning LoRA 문제 미해결) - 4-5: 한옥 summer (green pines) → 17.7%/3.97% 학습: - Redux strength 0.5 → 0.7 → 0.85 → 1.0 단조 증가 (56.41% sweet spot) - PuLID 0.7 fixed, Redux 1.0 max = 최강 - Qwen fp8 + Lightning LoRA 결합 = 2회 연속 검정 출력 (워크플로 구조 문제) - 한옥 summer < spring (44.45%) — 색감 자체가 차분 진행: R9 legacy 44.5% → R1 48.5% → R2 49.99% → R3 50.47% → R4 56.41% (+11.91pp) 자산 라이브러리: - workflows: winner_auto 7 + partial 7 - prompts/failed: 9 패턴 (Qwen 계열 2건 추가)
2026-05-19 23:49:55 +09:00
<div class="label-meta">models=flux-dev, pulid-flux</div>
<div class="label-auto-eval">pastel=25.65% / black=1.95%</div>
meeting-103 v2.1 clean R1 ✅5/5: PuLID + Redux = pastel 48.5% (R9 best 갱신) R0 환경 점검: - VRAM 11.7/12.4GB free (94%) — 해소 확인 - pytorch 2.12.0+cu130 + sageattention venv 검증 - 모델 인벤토리: Flux Dev Q5 GGUF, PuLID, Redux, byeolyi-ref-916 확보 - fp8 다운로드 .part 0 bytes (재개 대기) R1 clean 5조합 (5/5 success): - 1-1: PuLID w0.85 mean → pastel 25.7% / black 1.95% - 1-2: PuLID w0.85 mean delayed_start 0.05 → pastel 10.0% / black 14.0% (어두움) - 1-3: PuLID w1.0 mean → pastel 27.5% / black 0.79% - 1-4: PuLID w0.7 + Redux 0.5 → pastel 48.5% / black 0.27% ★ R1 best - 1-5: 한옥 baseline Flux Dev 768² → pastel 12.9% / black 11.7% CUDA error 회복: - 1-2 retry concat → CUDA invalid argument → 후속 1-2~1-5 모두 실패 - ComfyUI venv 재시작 (sage + pytorch 2.12) → 1-2 변형 (delayed_start)로 대체 - 재시도 4/4 모두 success 이전 R9 best (44.5%) → R1-4 (48.5%) = +4 pp 갱신 자산 라이브러리 누적: - models-performance.json: 7 모델 추적 (flux-dev, pulid-flux, flux-redux 등) - workflows: winner_auto 1 + partial 2 - prompts/proven: illustration_auto 패턴 3건 - prompts/failed: pastel-low + CUDA error 패턴 - prompts/negative: A_byeolyi 패턴 etc. 검수: - Vault: D:\Vault\8460s-image-rd\eval\meeting-103-v2.1-clean-grid.html - Gitea: results/meeting-103-v2.1-clean/round1/
2026-05-19 23:17:19 +09:00
<div class="label-flags"><span class="flag flag-ok">clean</span></div>
</div>
</div>
<div class="cell status-success">
<img src="../results/meeting-103-v2.1-clean/round1/1-2_thumb.png" data-original="../results/meeting-103-v2.1-clean/round1/1-2.png" alt="1-2">
<div class="label">
meeting-103 v2.1 clean R2 ✅5/5: PuLID 0.85+Redux 0.5 = pastel 49.99% (전체 best 갱신) R2 5조합 (5/5 success, retry 2회 포함): - 2-1: PuLID 0.7 + Redux 0.5 + Stage 3 Refiner (denoise 0.3) → 48.82% (Stage 3 효과 미미) - 2-2: PuLID 0.7 + Redux 0.7 → 48.93% - 2-3: PuLID 0.7 + Redux 0.3 → 38.52% (Redux 낮으면 파스텔 감소) - 2-4: PuLID 0.85 + Redux 0.5 → 49.99% ★ R2 best & 전체 best - 2-5: 한옥 spring (cherry blossom) → 44.45% / black 4.3% R1 best (R1-4 48.5%) → R2-4 (49.99%) = +1.5 pp 갱신 R9 legacy (44.5%) → R2-4 (49.99%) = +5.5 pp 누적 학습: - PuLID weight 0.85 + Redux 0.5 multiply 조합이 0.7/0.5 보다 우월 - Redux 0.3 너무 낮음, 0.7 비슷, 0.85+ 미검증 - Stage 3 Refiner 효과 미미 (48.5% → 48.82%, +0.3pp) CUDA error 회복: - 2-3, 2-4, 2-5 첫 시도 CUDA invalid argument - ComfyUI venv 재시작 후 retry → 3/3 모두 success - 매 ~5건 후 재시작 패턴 확인됨 자산 라이브러리: - models-performance: 8 모델 추적 - workflows: winner_auto 4 (R1-4, R2-1, R2-2, R2-4) + partial 5 - prompts/proven: 7 패턴 (R1, R2 합산) R3 계획: - 3-1: PuLID 0.7 + Redux 0.85 (push redux 더 강) - 3-2: PuLID 0.6 + Redux 0.7 (PuLID 약화) - 3-3: PuLID 0.7 + Redux 0.7 @ 1024x1792 (해상도 증가) - 3-4: Qwen-Image fp8 + Lightning LoRA 60갑자 (fp8 ~20G 완료, GGUF 회피) - 3-5: Pony V6 img2img from byeolyi-ref (대안 ID 방법) 검수: D:\Vault\8460s-image-rd\eval\meeting-103-v2.1-clean-grid.html
2026-05-19 23:36:44 +09:00
<div class="label-title">R1-1-2: A — PuLID weight 0.85 mean delayed_start 0.05 (concat-replacement)</div>
meeting-103 v2.1 clean R3+R4 ✅: PuLID 0.7+Redux 1.0 = pastel 56.41% (전체 best 갱신) R3 (4/5 success): - 3-1: PuLID 0.7 + Redux 0.85 → 50.47% (R2 best 49.99% 갱신) - 3-2: PuLID 0.6 + Redux 0.7 → 49.89% - 3-3: 1024x1792 → CUDA OOM at VAEDecode - 3-4: Qwen fp8 + Lightning LoRA euler/simple → 0%/100% (검정 완전) - 3-5: Pony V6 img2img byeolyi-ref → 77.3% (img2img 0.65 denoise, 원본 색 보존) R4 (3/5 success, 별이 새 best): - 4-1: PuLID 0.7 + Redux 1.0 (max) → 56.41% ★ 전체 best - 4-2: PuLID 0.85 + Redux 0.7 → 50.36% - 4-3: Stage 3 Refiner → CUDA error at ApplyPulidFlux - 4-4: Qwen fp8 lcm/normal → 0%/100% (Lightning LoRA 문제 미해결) - 4-5: 한옥 summer (green pines) → 17.7%/3.97% 학습: - Redux strength 0.5 → 0.7 → 0.85 → 1.0 단조 증가 (56.41% sweet spot) - PuLID 0.7 fixed, Redux 1.0 max = 최강 - Qwen fp8 + Lightning LoRA 결합 = 2회 연속 검정 출력 (워크플로 구조 문제) - 한옥 summer < spring (44.45%) — 색감 자체가 차분 진행: R9 legacy 44.5% → R1 48.5% → R2 49.99% → R3 50.47% → R4 56.41% (+11.91pp) 자산 라이브러리: - workflows: winner_auto 7 + partial 7 - prompts/failed: 9 패턴 (Qwen 계열 2건 추가)
2026-05-19 23:49:55 +09:00
<div class="label-meta">models=flux-dev, pulid-flux</div>
<div class="label-auto-eval">pastel=9.99% / black=13.98%</div>
meeting-103 v2.1 clean R2 ✅5/5: PuLID 0.85+Redux 0.5 = pastel 49.99% (전체 best 갱신) R2 5조합 (5/5 success, retry 2회 포함): - 2-1: PuLID 0.7 + Redux 0.5 + Stage 3 Refiner (denoise 0.3) → 48.82% (Stage 3 효과 미미) - 2-2: PuLID 0.7 + Redux 0.7 → 48.93% - 2-3: PuLID 0.7 + Redux 0.3 → 38.52% (Redux 낮으면 파스텔 감소) - 2-4: PuLID 0.85 + Redux 0.5 → 49.99% ★ R2 best & 전체 best - 2-5: 한옥 spring (cherry blossom) → 44.45% / black 4.3% R1 best (R1-4 48.5%) → R2-4 (49.99%) = +1.5 pp 갱신 R9 legacy (44.5%) → R2-4 (49.99%) = +5.5 pp 누적 학습: - PuLID weight 0.85 + Redux 0.5 multiply 조합이 0.7/0.5 보다 우월 - Redux 0.3 너무 낮음, 0.7 비슷, 0.85+ 미검증 - Stage 3 Refiner 효과 미미 (48.5% → 48.82%, +0.3pp) CUDA error 회복: - 2-3, 2-4, 2-5 첫 시도 CUDA invalid argument - ComfyUI venv 재시작 후 retry → 3/3 모두 success - 매 ~5건 후 재시작 패턴 확인됨 자산 라이브러리: - models-performance: 8 모델 추적 - workflows: winner_auto 4 (R1-4, R2-1, R2-2, R2-4) + partial 5 - prompts/proven: 7 패턴 (R1, R2 합산) R3 계획: - 3-1: PuLID 0.7 + Redux 0.85 (push redux 더 강) - 3-2: PuLID 0.6 + Redux 0.7 (PuLID 약화) - 3-3: PuLID 0.7 + Redux 0.7 @ 1024x1792 (해상도 증가) - 3-4: Qwen-Image fp8 + Lightning LoRA 60갑자 (fp8 ~20G 완료, GGUF 회피) - 3-5: Pony V6 img2img from byeolyi-ref (대안 ID 방법) 검수: D:\Vault\8460s-image-rd\eval\meeting-103-v2.1-clean-grid.html
2026-05-19 23:36:44 +09:00
<div class="label-flags"><span class="flag flag-warn">black-suspicious</span><span class="flag flag-warn">pastel-low</span></div>
meeting-103 v2.1 clean R1 ✅5/5: PuLID + Redux = pastel 48.5% (R9 best 갱신) R0 환경 점검: - VRAM 11.7/12.4GB free (94%) — 해소 확인 - pytorch 2.12.0+cu130 + sageattention venv 검증 - 모델 인벤토리: Flux Dev Q5 GGUF, PuLID, Redux, byeolyi-ref-916 확보 - fp8 다운로드 .part 0 bytes (재개 대기) R1 clean 5조합 (5/5 success): - 1-1: PuLID w0.85 mean → pastel 25.7% / black 1.95% - 1-2: PuLID w0.85 mean delayed_start 0.05 → pastel 10.0% / black 14.0% (어두움) - 1-3: PuLID w1.0 mean → pastel 27.5% / black 0.79% - 1-4: PuLID w0.7 + Redux 0.5 → pastel 48.5% / black 0.27% ★ R1 best - 1-5: 한옥 baseline Flux Dev 768² → pastel 12.9% / black 11.7% CUDA error 회복: - 1-2 retry concat → CUDA invalid argument → 후속 1-2~1-5 모두 실패 - ComfyUI venv 재시작 (sage + pytorch 2.12) → 1-2 변형 (delayed_start)로 대체 - 재시도 4/4 모두 success 이전 R9 best (44.5%) → R1-4 (48.5%) = +4 pp 갱신 자산 라이브러리 누적: - models-performance.json: 7 모델 추적 (flux-dev, pulid-flux, flux-redux 등) - workflows: winner_auto 1 + partial 2 - prompts/proven: illustration_auto 패턴 3건 - prompts/failed: pastel-low + CUDA error 패턴 - prompts/negative: A_byeolyi 패턴 etc. 검수: - Vault: D:\Vault\8460s-image-rd\eval\meeting-103-v2.1-clean-grid.html - Gitea: results/meeting-103-v2.1-clean/round1/
2026-05-19 23:17:19 +09:00
</div>
</div>
<div class="cell status-success">
<img src="../results/meeting-103-v2.1-clean/round1/1-3_thumb.png" data-original="../results/meeting-103-v2.1-clean/round1/1-3.png" alt="1-3">
<div class="label">
meeting-103 v2.1 clean R2 ✅5/5: PuLID 0.85+Redux 0.5 = pastel 49.99% (전체 best 갱신) R2 5조합 (5/5 success, retry 2회 포함): - 2-1: PuLID 0.7 + Redux 0.5 + Stage 3 Refiner (denoise 0.3) → 48.82% (Stage 3 효과 미미) - 2-2: PuLID 0.7 + Redux 0.7 → 48.93% - 2-3: PuLID 0.7 + Redux 0.3 → 38.52% (Redux 낮으면 파스텔 감소) - 2-4: PuLID 0.85 + Redux 0.5 → 49.99% ★ R2 best & 전체 best - 2-5: 한옥 spring (cherry blossom) → 44.45% / black 4.3% R1 best (R1-4 48.5%) → R2-4 (49.99%) = +1.5 pp 갱신 R9 legacy (44.5%) → R2-4 (49.99%) = +5.5 pp 누적 학습: - PuLID weight 0.85 + Redux 0.5 multiply 조합이 0.7/0.5 보다 우월 - Redux 0.3 너무 낮음, 0.7 비슷, 0.85+ 미검증 - Stage 3 Refiner 효과 미미 (48.5% → 48.82%, +0.3pp) CUDA error 회복: - 2-3, 2-4, 2-5 첫 시도 CUDA invalid argument - ComfyUI venv 재시작 후 retry → 3/3 모두 success - 매 ~5건 후 재시작 패턴 확인됨 자산 라이브러리: - models-performance: 8 모델 추적 - workflows: winner_auto 4 (R1-4, R2-1, R2-2, R2-4) + partial 5 - prompts/proven: 7 패턴 (R1, R2 합산) R3 계획: - 3-1: PuLID 0.7 + Redux 0.85 (push redux 더 강) - 3-2: PuLID 0.6 + Redux 0.7 (PuLID 약화) - 3-3: PuLID 0.7 + Redux 0.7 @ 1024x1792 (해상도 증가) - 3-4: Qwen-Image fp8 + Lightning LoRA 60갑자 (fp8 ~20G 완료, GGUF 회피) - 3-5: Pony V6 img2img from byeolyi-ref (대안 ID 방법) 검수: D:\Vault\8460s-image-rd\eval\meeting-103-v2.1-clean-grid.html
2026-05-19 23:36:44 +09:00
<div class="label-title">R1-1-3: A — PuLID weight 1.0 mean</div>
meeting-103 v2.1 clean R3+R4 ✅: PuLID 0.7+Redux 1.0 = pastel 56.41% (전체 best 갱신) R3 (4/5 success): - 3-1: PuLID 0.7 + Redux 0.85 → 50.47% (R2 best 49.99% 갱신) - 3-2: PuLID 0.6 + Redux 0.7 → 49.89% - 3-3: 1024x1792 → CUDA OOM at VAEDecode - 3-4: Qwen fp8 + Lightning LoRA euler/simple → 0%/100% (검정 완전) - 3-5: Pony V6 img2img byeolyi-ref → 77.3% (img2img 0.65 denoise, 원본 색 보존) R4 (3/5 success, 별이 새 best): - 4-1: PuLID 0.7 + Redux 1.0 (max) → 56.41% ★ 전체 best - 4-2: PuLID 0.85 + Redux 0.7 → 50.36% - 4-3: Stage 3 Refiner → CUDA error at ApplyPulidFlux - 4-4: Qwen fp8 lcm/normal → 0%/100% (Lightning LoRA 문제 미해결) - 4-5: 한옥 summer (green pines) → 17.7%/3.97% 학습: - Redux strength 0.5 → 0.7 → 0.85 → 1.0 단조 증가 (56.41% sweet spot) - PuLID 0.7 fixed, Redux 1.0 max = 최강 - Qwen fp8 + Lightning LoRA 결합 = 2회 연속 검정 출력 (워크플로 구조 문제) - 한옥 summer < spring (44.45%) — 색감 자체가 차분 진행: R9 legacy 44.5% → R1 48.5% → R2 49.99% → R3 50.47% → R4 56.41% (+11.91pp) 자산 라이브러리: - workflows: winner_auto 7 + partial 7 - prompts/failed: 9 패턴 (Qwen 계열 2건 추가)
2026-05-19 23:49:55 +09:00
<div class="label-meta">models=flux-dev, pulid-flux</div>
<div class="label-auto-eval">pastel=27.51% / black=0.79%</div>
meeting-103 v2.1 clean R1 ✅5/5: PuLID + Redux = pastel 48.5% (R9 best 갱신) R0 환경 점검: - VRAM 11.7/12.4GB free (94%) — 해소 확인 - pytorch 2.12.0+cu130 + sageattention venv 검증 - 모델 인벤토리: Flux Dev Q5 GGUF, PuLID, Redux, byeolyi-ref-916 확보 - fp8 다운로드 .part 0 bytes (재개 대기) R1 clean 5조합 (5/5 success): - 1-1: PuLID w0.85 mean → pastel 25.7% / black 1.95% - 1-2: PuLID w0.85 mean delayed_start 0.05 → pastel 10.0% / black 14.0% (어두움) - 1-3: PuLID w1.0 mean → pastel 27.5% / black 0.79% - 1-4: PuLID w0.7 + Redux 0.5 → pastel 48.5% / black 0.27% ★ R1 best - 1-5: 한옥 baseline Flux Dev 768² → pastel 12.9% / black 11.7% CUDA error 회복: - 1-2 retry concat → CUDA invalid argument → 후속 1-2~1-5 모두 실패 - ComfyUI venv 재시작 (sage + pytorch 2.12) → 1-2 변형 (delayed_start)로 대체 - 재시도 4/4 모두 success 이전 R9 best (44.5%) → R1-4 (48.5%) = +4 pp 갱신 자산 라이브러리 누적: - models-performance.json: 7 모델 추적 (flux-dev, pulid-flux, flux-redux 등) - workflows: winner_auto 1 + partial 2 - prompts/proven: illustration_auto 패턴 3건 - prompts/failed: pastel-low + CUDA error 패턴 - prompts/negative: A_byeolyi 패턴 etc. 검수: - Vault: D:\Vault\8460s-image-rd\eval\meeting-103-v2.1-clean-grid.html - Gitea: results/meeting-103-v2.1-clean/round1/
2026-05-19 23:17:19 +09:00
<div class="label-flags"><span class="flag flag-ok">clean</span></div>
</div>
</div>
<div class="cell status-success">
<img src="../results/meeting-103-v2.1-clean/round1/1-4_thumb.png" data-original="../results/meeting-103-v2.1-clean/round1/1-4.png" alt="1-4">
<div class="label">
meeting-103 v2.1 clean R2 ✅5/5: PuLID 0.85+Redux 0.5 = pastel 49.99% (전체 best 갱신) R2 5조합 (5/5 success, retry 2회 포함): - 2-1: PuLID 0.7 + Redux 0.5 + Stage 3 Refiner (denoise 0.3) → 48.82% (Stage 3 효과 미미) - 2-2: PuLID 0.7 + Redux 0.7 → 48.93% - 2-3: PuLID 0.7 + Redux 0.3 → 38.52% (Redux 낮으면 파스텔 감소) - 2-4: PuLID 0.85 + Redux 0.5 → 49.99% ★ R2 best & 전체 best - 2-5: 한옥 spring (cherry blossom) → 44.45% / black 4.3% R1 best (R1-4 48.5%) → R2-4 (49.99%) = +1.5 pp 갱신 R9 legacy (44.5%) → R2-4 (49.99%) = +5.5 pp 누적 학습: - PuLID weight 0.85 + Redux 0.5 multiply 조합이 0.7/0.5 보다 우월 - Redux 0.3 너무 낮음, 0.7 비슷, 0.85+ 미검증 - Stage 3 Refiner 효과 미미 (48.5% → 48.82%, +0.3pp) CUDA error 회복: - 2-3, 2-4, 2-5 첫 시도 CUDA invalid argument - ComfyUI venv 재시작 후 retry → 3/3 모두 success - 매 ~5건 후 재시작 패턴 확인됨 자산 라이브러리: - models-performance: 8 모델 추적 - workflows: winner_auto 4 (R1-4, R2-1, R2-2, R2-4) + partial 5 - prompts/proven: 7 패턴 (R1, R2 합산) R3 계획: - 3-1: PuLID 0.7 + Redux 0.85 (push redux 더 강) - 3-2: PuLID 0.6 + Redux 0.7 (PuLID 약화) - 3-3: PuLID 0.7 + Redux 0.7 @ 1024x1792 (해상도 증가) - 3-4: Qwen-Image fp8 + Lightning LoRA 60갑자 (fp8 ~20G 완료, GGUF 회피) - 3-5: Pony V6 img2img from byeolyi-ref (대안 ID 방법) 검수: D:\Vault\8460s-image-rd\eval\meeting-103-v2.1-clean-grid.html
2026-05-19 23:36:44 +09:00
<div class="label-title">R1-1-4: A — PuLID 0.7 + Redux 0.5</div>
meeting-103 v2.1 clean R3+R4 ✅: PuLID 0.7+Redux 1.0 = pastel 56.41% (전체 best 갱신) R3 (4/5 success): - 3-1: PuLID 0.7 + Redux 0.85 → 50.47% (R2 best 49.99% 갱신) - 3-2: PuLID 0.6 + Redux 0.7 → 49.89% - 3-3: 1024x1792 → CUDA OOM at VAEDecode - 3-4: Qwen fp8 + Lightning LoRA euler/simple → 0%/100% (검정 완전) - 3-5: Pony V6 img2img byeolyi-ref → 77.3% (img2img 0.65 denoise, 원본 색 보존) R4 (3/5 success, 별이 새 best): - 4-1: PuLID 0.7 + Redux 1.0 (max) → 56.41% ★ 전체 best - 4-2: PuLID 0.85 + Redux 0.7 → 50.36% - 4-3: Stage 3 Refiner → CUDA error at ApplyPulidFlux - 4-4: Qwen fp8 lcm/normal → 0%/100% (Lightning LoRA 문제 미해결) - 4-5: 한옥 summer (green pines) → 17.7%/3.97% 학습: - Redux strength 0.5 → 0.7 → 0.85 → 1.0 단조 증가 (56.41% sweet spot) - PuLID 0.7 fixed, Redux 1.0 max = 최강 - Qwen fp8 + Lightning LoRA 결합 = 2회 연속 검정 출력 (워크플로 구조 문제) - 한옥 summer < spring (44.45%) — 색감 자체가 차분 진행: R9 legacy 44.5% → R1 48.5% → R2 49.99% → R3 50.47% → R4 56.41% (+11.91pp) 자산 라이브러리: - workflows: winner_auto 7 + partial 7 - prompts/failed: 9 패턴 (Qwen 계열 2건 추가)
2026-05-19 23:49:55 +09:00
<div class="label-meta">models=flux-dev, pulid-flux, flux-redux</div>
<div class="label-auto-eval">pastel=48.5% / black=0.27%</div>
meeting-103 v2.1 clean R1 ✅5/5: PuLID + Redux = pastel 48.5% (R9 best 갱신) R0 환경 점검: - VRAM 11.7/12.4GB free (94%) — 해소 확인 - pytorch 2.12.0+cu130 + sageattention venv 검증 - 모델 인벤토리: Flux Dev Q5 GGUF, PuLID, Redux, byeolyi-ref-916 확보 - fp8 다운로드 .part 0 bytes (재개 대기) R1 clean 5조합 (5/5 success): - 1-1: PuLID w0.85 mean → pastel 25.7% / black 1.95% - 1-2: PuLID w0.85 mean delayed_start 0.05 → pastel 10.0% / black 14.0% (어두움) - 1-3: PuLID w1.0 mean → pastel 27.5% / black 0.79% - 1-4: PuLID w0.7 + Redux 0.5 → pastel 48.5% / black 0.27% ★ R1 best - 1-5: 한옥 baseline Flux Dev 768² → pastel 12.9% / black 11.7% CUDA error 회복: - 1-2 retry concat → CUDA invalid argument → 후속 1-2~1-5 모두 실패 - ComfyUI venv 재시작 (sage + pytorch 2.12) → 1-2 변형 (delayed_start)로 대체 - 재시도 4/4 모두 success 이전 R9 best (44.5%) → R1-4 (48.5%) = +4 pp 갱신 자산 라이브러리 누적: - models-performance.json: 7 모델 추적 (flux-dev, pulid-flux, flux-redux 등) - workflows: winner_auto 1 + partial 2 - prompts/proven: illustration_auto 패턴 3건 - prompts/failed: pastel-low + CUDA error 패턴 - prompts/negative: A_byeolyi 패턴 etc. 검수: - Vault: D:\Vault\8460s-image-rd\eval\meeting-103-v2.1-clean-grid.html - Gitea: results/meeting-103-v2.1-clean/round1/
2026-05-19 23:17:19 +09:00
<div class="label-flags"><span class="flag flag-ok">clean</span></div>
</div>
</div>
<div class="cell status-success">
<img src="../results/meeting-103-v2.1-clean/round1/1-5_thumb.png" data-original="../results/meeting-103-v2.1-clean/round1/1-5.png" alt="1-5">
<div class="label">
meeting-103 v2.1 clean R2 ✅5/5: PuLID 0.85+Redux 0.5 = pastel 49.99% (전체 best 갱신) R2 5조합 (5/5 success, retry 2회 포함): - 2-1: PuLID 0.7 + Redux 0.5 + Stage 3 Refiner (denoise 0.3) → 48.82% (Stage 3 효과 미미) - 2-2: PuLID 0.7 + Redux 0.7 → 48.93% - 2-3: PuLID 0.7 + Redux 0.3 → 38.52% (Redux 낮으면 파스텔 감소) - 2-4: PuLID 0.85 + Redux 0.5 → 49.99% ★ R2 best & 전체 best - 2-5: 한옥 spring (cherry blossom) → 44.45% / black 4.3% R1 best (R1-4 48.5%) → R2-4 (49.99%) = +1.5 pp 갱신 R9 legacy (44.5%) → R2-4 (49.99%) = +5.5 pp 누적 학습: - PuLID weight 0.85 + Redux 0.5 multiply 조합이 0.7/0.5 보다 우월 - Redux 0.3 너무 낮음, 0.7 비슷, 0.85+ 미검증 - Stage 3 Refiner 효과 미미 (48.5% → 48.82%, +0.3pp) CUDA error 회복: - 2-3, 2-4, 2-5 첫 시도 CUDA invalid argument - ComfyUI venv 재시작 후 retry → 3/3 모두 success - 매 ~5건 후 재시작 패턴 확인됨 자산 라이브러리: - models-performance: 8 모델 추적 - workflows: winner_auto 4 (R1-4, R2-1, R2-2, R2-4) + partial 5 - prompts/proven: 7 패턴 (R1, R2 합산) R3 계획: - 3-1: PuLID 0.7 + Redux 0.85 (push redux 더 강) - 3-2: PuLID 0.6 + Redux 0.7 (PuLID 약화) - 3-3: PuLID 0.7 + Redux 0.7 @ 1024x1792 (해상도 증가) - 3-4: Qwen-Image fp8 + Lightning LoRA 60갑자 (fp8 ~20G 완료, GGUF 회피) - 3-5: Pony V6 img2img from byeolyi-ref (대안 ID 방법) 검수: D:\Vault\8460s-image-rd\eval\meeting-103-v2.1-clean-grid.html
2026-05-19 23:36:44 +09:00
<div class="label-title">R1-1-5: B — Hanok baseline Flux Dev 768²</div>
meeting-103 v2.1 clean R3+R4 ✅: PuLID 0.7+Redux 1.0 = pastel 56.41% (전체 best 갱신) R3 (4/5 success): - 3-1: PuLID 0.7 + Redux 0.85 → 50.47% (R2 best 49.99% 갱신) - 3-2: PuLID 0.6 + Redux 0.7 → 49.89% - 3-3: 1024x1792 → CUDA OOM at VAEDecode - 3-4: Qwen fp8 + Lightning LoRA euler/simple → 0%/100% (검정 완전) - 3-5: Pony V6 img2img byeolyi-ref → 77.3% (img2img 0.65 denoise, 원본 색 보존) R4 (3/5 success, 별이 새 best): - 4-1: PuLID 0.7 + Redux 1.0 (max) → 56.41% ★ 전체 best - 4-2: PuLID 0.85 + Redux 0.7 → 50.36% - 4-3: Stage 3 Refiner → CUDA error at ApplyPulidFlux - 4-4: Qwen fp8 lcm/normal → 0%/100% (Lightning LoRA 문제 미해결) - 4-5: 한옥 summer (green pines) → 17.7%/3.97% 학습: - Redux strength 0.5 → 0.7 → 0.85 → 1.0 단조 증가 (56.41% sweet spot) - PuLID 0.7 fixed, Redux 1.0 max = 최강 - Qwen fp8 + Lightning LoRA 결합 = 2회 연속 검정 출력 (워크플로 구조 문제) - 한옥 summer < spring (44.45%) — 색감 자체가 차분 진행: R9 legacy 44.5% → R1 48.5% → R2 49.99% → R3 50.47% → R4 56.41% (+11.91pp) 자산 라이브러리: - workflows: winner_auto 7 + partial 7 - prompts/failed: 9 패턴 (Qwen 계열 2건 추가)
2026-05-19 23:49:55 +09:00
<div class="label-meta">models=flux-dev</div>
<div class="label-auto-eval">pastel=12.87% / black=11.67%</div>
meeting-103 v2.1 clean R2 ✅5/5: PuLID 0.85+Redux 0.5 = pastel 49.99% (전체 best 갱신) R2 5조합 (5/5 success, retry 2회 포함): - 2-1: PuLID 0.7 + Redux 0.5 + Stage 3 Refiner (denoise 0.3) → 48.82% (Stage 3 효과 미미) - 2-2: PuLID 0.7 + Redux 0.7 → 48.93% - 2-3: PuLID 0.7 + Redux 0.3 → 38.52% (Redux 낮으면 파스텔 감소) - 2-4: PuLID 0.85 + Redux 0.5 → 49.99% ★ R2 best & 전체 best - 2-5: 한옥 spring (cherry blossom) → 44.45% / black 4.3% R1 best (R1-4 48.5%) → R2-4 (49.99%) = +1.5 pp 갱신 R9 legacy (44.5%) → R2-4 (49.99%) = +5.5 pp 누적 학습: - PuLID weight 0.85 + Redux 0.5 multiply 조합이 0.7/0.5 보다 우월 - Redux 0.3 너무 낮음, 0.7 비슷, 0.85+ 미검증 - Stage 3 Refiner 효과 미미 (48.5% → 48.82%, +0.3pp) CUDA error 회복: - 2-3, 2-4, 2-5 첫 시도 CUDA invalid argument - ComfyUI venv 재시작 후 retry → 3/3 모두 success - 매 ~5건 후 재시작 패턴 확인됨 자산 라이브러리: - models-performance: 8 모델 추적 - workflows: winner_auto 4 (R1-4, R2-1, R2-2, R2-4) + partial 5 - prompts/proven: 7 패턴 (R1, R2 합산) R3 계획: - 3-1: PuLID 0.7 + Redux 0.85 (push redux 더 강) - 3-2: PuLID 0.6 + Redux 0.7 (PuLID 약화) - 3-3: PuLID 0.7 + Redux 0.7 @ 1024x1792 (해상도 증가) - 3-4: Qwen-Image fp8 + Lightning LoRA 60갑자 (fp8 ~20G 완료, GGUF 회피) - 3-5: Pony V6 img2img from byeolyi-ref (대안 ID 방법) 검수: D:\Vault\8460s-image-rd\eval\meeting-103-v2.1-clean-grid.html
2026-05-19 23:36:44 +09:00
<div class="label-flags"><span class="flag flag-warn">black-suspicious</span></div>
</div>
</div>
<div class="cell status-success">
<img src="../results/meeting-103-v2.1-clean/round2/2-1_thumb.png" data-original="../results/meeting-103-v2.1-clean/round2/2-1.png" alt="2-1">
<div class="label">
<div class="label-title">R2-2-1: A — PuLID 0.7 + Redux 0.5 + Stage 3 Refiner (denoise 0.3)</div>
meeting-103 v2.1 clean R3+R4 ✅: PuLID 0.7+Redux 1.0 = pastel 56.41% (전체 best 갱신) R3 (4/5 success): - 3-1: PuLID 0.7 + Redux 0.85 → 50.47% (R2 best 49.99% 갱신) - 3-2: PuLID 0.6 + Redux 0.7 → 49.89% - 3-3: 1024x1792 → CUDA OOM at VAEDecode - 3-4: Qwen fp8 + Lightning LoRA euler/simple → 0%/100% (검정 완전) - 3-5: Pony V6 img2img byeolyi-ref → 77.3% (img2img 0.65 denoise, 원본 색 보존) R4 (3/5 success, 별이 새 best): - 4-1: PuLID 0.7 + Redux 1.0 (max) → 56.41% ★ 전체 best - 4-2: PuLID 0.85 + Redux 0.7 → 50.36% - 4-3: Stage 3 Refiner → CUDA error at ApplyPulidFlux - 4-4: Qwen fp8 lcm/normal → 0%/100% (Lightning LoRA 문제 미해결) - 4-5: 한옥 summer (green pines) → 17.7%/3.97% 학습: - Redux strength 0.5 → 0.7 → 0.85 → 1.0 단조 증가 (56.41% sweet spot) - PuLID 0.7 fixed, Redux 1.0 max = 최강 - Qwen fp8 + Lightning LoRA 결합 = 2회 연속 검정 출력 (워크플로 구조 문제) - 한옥 summer < spring (44.45%) — 색감 자체가 차분 진행: R9 legacy 44.5% → R1 48.5% → R2 49.99% → R3 50.47% → R4 56.41% (+11.91pp) 자산 라이브러리: - workflows: winner_auto 7 + partial 7 - prompts/failed: 9 패턴 (Qwen 계열 2건 추가)
2026-05-19 23:49:55 +09:00
<div class="label-meta">models=flux-dev, pulid-flux, flux-redux, flux-dev-refiner</div>
<div class="label-auto-eval">pastel=48.82% / black=0.34%</div>
meeting-103 v2.1 clean R2 ✅5/5: PuLID 0.85+Redux 0.5 = pastel 49.99% (전체 best 갱신) R2 5조합 (5/5 success, retry 2회 포함): - 2-1: PuLID 0.7 + Redux 0.5 + Stage 3 Refiner (denoise 0.3) → 48.82% (Stage 3 효과 미미) - 2-2: PuLID 0.7 + Redux 0.7 → 48.93% - 2-3: PuLID 0.7 + Redux 0.3 → 38.52% (Redux 낮으면 파스텔 감소) - 2-4: PuLID 0.85 + Redux 0.5 → 49.99% ★ R2 best & 전체 best - 2-5: 한옥 spring (cherry blossom) → 44.45% / black 4.3% R1 best (R1-4 48.5%) → R2-4 (49.99%) = +1.5 pp 갱신 R9 legacy (44.5%) → R2-4 (49.99%) = +5.5 pp 누적 학습: - PuLID weight 0.85 + Redux 0.5 multiply 조합이 0.7/0.5 보다 우월 - Redux 0.3 너무 낮음, 0.7 비슷, 0.85+ 미검증 - Stage 3 Refiner 효과 미미 (48.5% → 48.82%, +0.3pp) CUDA error 회복: - 2-3, 2-4, 2-5 첫 시도 CUDA invalid argument - ComfyUI venv 재시작 후 retry → 3/3 모두 success - 매 ~5건 후 재시작 패턴 확인됨 자산 라이브러리: - models-performance: 8 모델 추적 - workflows: winner_auto 4 (R1-4, R2-1, R2-2, R2-4) + partial 5 - prompts/proven: 7 패턴 (R1, R2 합산) R3 계획: - 3-1: PuLID 0.7 + Redux 0.85 (push redux 더 강) - 3-2: PuLID 0.6 + Redux 0.7 (PuLID 약화) - 3-3: PuLID 0.7 + Redux 0.7 @ 1024x1792 (해상도 증가) - 3-4: Qwen-Image fp8 + Lightning LoRA 60갑자 (fp8 ~20G 완료, GGUF 회피) - 3-5: Pony V6 img2img from byeolyi-ref (대안 ID 방법) 검수: D:\Vault\8460s-image-rd\eval\meeting-103-v2.1-clean-grid.html
2026-05-19 23:36:44 +09:00
<div class="label-flags"><span class="flag flag-ok">clean</span></div>
</div>
</div>
<div class="cell status-success">
<img src="../results/meeting-103-v2.1-clean/round2/2-2_thumb.png" data-original="../results/meeting-103-v2.1-clean/round2/2-2.png" alt="2-2">
<div class="label">
<div class="label-title">R2-2-2: A — PuLID 0.7 + Redux 0.7</div>
meeting-103 v2.1 clean R3+R4 ✅: PuLID 0.7+Redux 1.0 = pastel 56.41% (전체 best 갱신) R3 (4/5 success): - 3-1: PuLID 0.7 + Redux 0.85 → 50.47% (R2 best 49.99% 갱신) - 3-2: PuLID 0.6 + Redux 0.7 → 49.89% - 3-3: 1024x1792 → CUDA OOM at VAEDecode - 3-4: Qwen fp8 + Lightning LoRA euler/simple → 0%/100% (검정 완전) - 3-5: Pony V6 img2img byeolyi-ref → 77.3% (img2img 0.65 denoise, 원본 색 보존) R4 (3/5 success, 별이 새 best): - 4-1: PuLID 0.7 + Redux 1.0 (max) → 56.41% ★ 전체 best - 4-2: PuLID 0.85 + Redux 0.7 → 50.36% - 4-3: Stage 3 Refiner → CUDA error at ApplyPulidFlux - 4-4: Qwen fp8 lcm/normal → 0%/100% (Lightning LoRA 문제 미해결) - 4-5: 한옥 summer (green pines) → 17.7%/3.97% 학습: - Redux strength 0.5 → 0.7 → 0.85 → 1.0 단조 증가 (56.41% sweet spot) - PuLID 0.7 fixed, Redux 1.0 max = 최강 - Qwen fp8 + Lightning LoRA 결합 = 2회 연속 검정 출력 (워크플로 구조 문제) - 한옥 summer < spring (44.45%) — 색감 자체가 차분 진행: R9 legacy 44.5% → R1 48.5% → R2 49.99% → R3 50.47% → R4 56.41% (+11.91pp) 자산 라이브러리: - workflows: winner_auto 7 + partial 7 - prompts/failed: 9 패턴 (Qwen 계열 2건 추가)
2026-05-19 23:49:55 +09:00
<div class="label-meta">models=flux-dev, pulid-flux, flux-redux</div>
<div class="label-auto-eval">pastel=48.93% / black=0.46%</div>
meeting-103 v2.1 clean R2 ✅5/5: PuLID 0.85+Redux 0.5 = pastel 49.99% (전체 best 갱신) R2 5조합 (5/5 success, retry 2회 포함): - 2-1: PuLID 0.7 + Redux 0.5 + Stage 3 Refiner (denoise 0.3) → 48.82% (Stage 3 효과 미미) - 2-2: PuLID 0.7 + Redux 0.7 → 48.93% - 2-3: PuLID 0.7 + Redux 0.3 → 38.52% (Redux 낮으면 파스텔 감소) - 2-4: PuLID 0.85 + Redux 0.5 → 49.99% ★ R2 best & 전체 best - 2-5: 한옥 spring (cherry blossom) → 44.45% / black 4.3% R1 best (R1-4 48.5%) → R2-4 (49.99%) = +1.5 pp 갱신 R9 legacy (44.5%) → R2-4 (49.99%) = +5.5 pp 누적 학습: - PuLID weight 0.85 + Redux 0.5 multiply 조합이 0.7/0.5 보다 우월 - Redux 0.3 너무 낮음, 0.7 비슷, 0.85+ 미검증 - Stage 3 Refiner 효과 미미 (48.5% → 48.82%, +0.3pp) CUDA error 회복: - 2-3, 2-4, 2-5 첫 시도 CUDA invalid argument - ComfyUI venv 재시작 후 retry → 3/3 모두 success - 매 ~5건 후 재시작 패턴 확인됨 자산 라이브러리: - models-performance: 8 모델 추적 - workflows: winner_auto 4 (R1-4, R2-1, R2-2, R2-4) + partial 5 - prompts/proven: 7 패턴 (R1, R2 합산) R3 계획: - 3-1: PuLID 0.7 + Redux 0.85 (push redux 더 강) - 3-2: PuLID 0.6 + Redux 0.7 (PuLID 약화) - 3-3: PuLID 0.7 + Redux 0.7 @ 1024x1792 (해상도 증가) - 3-4: Qwen-Image fp8 + Lightning LoRA 60갑자 (fp8 ~20G 완료, GGUF 회피) - 3-5: Pony V6 img2img from byeolyi-ref (대안 ID 방법) 검수: D:\Vault\8460s-image-rd\eval\meeting-103-v2.1-clean-grid.html
2026-05-19 23:36:44 +09:00
<div class="label-flags"><span class="flag flag-ok">clean</span></div>
</div>
</div>
<div class="cell status-success">
<img src="../results/meeting-103-v2.1-clean/round2/2-3_thumb.png" data-original="../results/meeting-103-v2.1-clean/round2/2-3.png" alt="2-3">
<div class="label">
<div class="label-title">R2-2-3: A — PuLID 0.7 + Redux 0.3</div>
meeting-103 v2.1 clean R3+R4 ✅: PuLID 0.7+Redux 1.0 = pastel 56.41% (전체 best 갱신) R3 (4/5 success): - 3-1: PuLID 0.7 + Redux 0.85 → 50.47% (R2 best 49.99% 갱신) - 3-2: PuLID 0.6 + Redux 0.7 → 49.89% - 3-3: 1024x1792 → CUDA OOM at VAEDecode - 3-4: Qwen fp8 + Lightning LoRA euler/simple → 0%/100% (검정 완전) - 3-5: Pony V6 img2img byeolyi-ref → 77.3% (img2img 0.65 denoise, 원본 색 보존) R4 (3/5 success, 별이 새 best): - 4-1: PuLID 0.7 + Redux 1.0 (max) → 56.41% ★ 전체 best - 4-2: PuLID 0.85 + Redux 0.7 → 50.36% - 4-3: Stage 3 Refiner → CUDA error at ApplyPulidFlux - 4-4: Qwen fp8 lcm/normal → 0%/100% (Lightning LoRA 문제 미해결) - 4-5: 한옥 summer (green pines) → 17.7%/3.97% 학습: - Redux strength 0.5 → 0.7 → 0.85 → 1.0 단조 증가 (56.41% sweet spot) - PuLID 0.7 fixed, Redux 1.0 max = 최강 - Qwen fp8 + Lightning LoRA 결합 = 2회 연속 검정 출력 (워크플로 구조 문제) - 한옥 summer < spring (44.45%) — 색감 자체가 차분 진행: R9 legacy 44.5% → R1 48.5% → R2 49.99% → R3 50.47% → R4 56.41% (+11.91pp) 자산 라이브러리: - workflows: winner_auto 7 + partial 7 - prompts/failed: 9 패턴 (Qwen 계열 2건 추가)
2026-05-19 23:49:55 +09:00
<div class="label-meta">models=flux-dev, pulid-flux, flux-redux</div>
<div class="label-auto-eval">pastel=38.52% / black=0.48%</div>
meeting-103 v2.1 clean R2 ✅5/5: PuLID 0.85+Redux 0.5 = pastel 49.99% (전체 best 갱신) R2 5조합 (5/5 success, retry 2회 포함): - 2-1: PuLID 0.7 + Redux 0.5 + Stage 3 Refiner (denoise 0.3) → 48.82% (Stage 3 효과 미미) - 2-2: PuLID 0.7 + Redux 0.7 → 48.93% - 2-3: PuLID 0.7 + Redux 0.3 → 38.52% (Redux 낮으면 파스텔 감소) - 2-4: PuLID 0.85 + Redux 0.5 → 49.99% ★ R2 best & 전체 best - 2-5: 한옥 spring (cherry blossom) → 44.45% / black 4.3% R1 best (R1-4 48.5%) → R2-4 (49.99%) = +1.5 pp 갱신 R9 legacy (44.5%) → R2-4 (49.99%) = +5.5 pp 누적 학습: - PuLID weight 0.85 + Redux 0.5 multiply 조합이 0.7/0.5 보다 우월 - Redux 0.3 너무 낮음, 0.7 비슷, 0.85+ 미검증 - Stage 3 Refiner 효과 미미 (48.5% → 48.82%, +0.3pp) CUDA error 회복: - 2-3, 2-4, 2-5 첫 시도 CUDA invalid argument - ComfyUI venv 재시작 후 retry → 3/3 모두 success - 매 ~5건 후 재시작 패턴 확인됨 자산 라이브러리: - models-performance: 8 모델 추적 - workflows: winner_auto 4 (R1-4, R2-1, R2-2, R2-4) + partial 5 - prompts/proven: 7 패턴 (R1, R2 합산) R3 계획: - 3-1: PuLID 0.7 + Redux 0.85 (push redux 더 강) - 3-2: PuLID 0.6 + Redux 0.7 (PuLID 약화) - 3-3: PuLID 0.7 + Redux 0.7 @ 1024x1792 (해상도 증가) - 3-4: Qwen-Image fp8 + Lightning LoRA 60갑자 (fp8 ~20G 완료, GGUF 회피) - 3-5: Pony V6 img2img from byeolyi-ref (대안 ID 방법) 검수: D:\Vault\8460s-image-rd\eval\meeting-103-v2.1-clean-grid.html
2026-05-19 23:36:44 +09:00
<div class="label-flags"><span class="flag flag-ok">clean</span></div>
</div>
</div>
<div class="cell status-success">
<img src="../results/meeting-103-v2.1-clean/round2/2-4_thumb.png" data-original="../results/meeting-103-v2.1-clean/round2/2-4.png" alt="2-4">
<div class="label">
<div class="label-title">R2-2-4: A — PuLID 0.85 + Redux 0.5</div>
meeting-103 v2.1 clean R3+R4 ✅: PuLID 0.7+Redux 1.0 = pastel 56.41% (전체 best 갱신) R3 (4/5 success): - 3-1: PuLID 0.7 + Redux 0.85 → 50.47% (R2 best 49.99% 갱신) - 3-2: PuLID 0.6 + Redux 0.7 → 49.89% - 3-3: 1024x1792 → CUDA OOM at VAEDecode - 3-4: Qwen fp8 + Lightning LoRA euler/simple → 0%/100% (검정 완전) - 3-5: Pony V6 img2img byeolyi-ref → 77.3% (img2img 0.65 denoise, 원본 색 보존) R4 (3/5 success, 별이 새 best): - 4-1: PuLID 0.7 + Redux 1.0 (max) → 56.41% ★ 전체 best - 4-2: PuLID 0.85 + Redux 0.7 → 50.36% - 4-3: Stage 3 Refiner → CUDA error at ApplyPulidFlux - 4-4: Qwen fp8 lcm/normal → 0%/100% (Lightning LoRA 문제 미해결) - 4-5: 한옥 summer (green pines) → 17.7%/3.97% 학습: - Redux strength 0.5 → 0.7 → 0.85 → 1.0 단조 증가 (56.41% sweet spot) - PuLID 0.7 fixed, Redux 1.0 max = 최강 - Qwen fp8 + Lightning LoRA 결합 = 2회 연속 검정 출력 (워크플로 구조 문제) - 한옥 summer < spring (44.45%) — 색감 자체가 차분 진행: R9 legacy 44.5% → R1 48.5% → R2 49.99% → R3 50.47% → R4 56.41% (+11.91pp) 자산 라이브러리: - workflows: winner_auto 7 + partial 7 - prompts/failed: 9 패턴 (Qwen 계열 2건 추가)
2026-05-19 23:49:55 +09:00
<div class="label-meta">models=flux-dev, pulid-flux, flux-redux</div>
<div class="label-auto-eval">pastel=49.99% / black=0.18%</div>
meeting-103 v2.1 clean R2 ✅5/5: PuLID 0.85+Redux 0.5 = pastel 49.99% (전체 best 갱신) R2 5조합 (5/5 success, retry 2회 포함): - 2-1: PuLID 0.7 + Redux 0.5 + Stage 3 Refiner (denoise 0.3) → 48.82% (Stage 3 효과 미미) - 2-2: PuLID 0.7 + Redux 0.7 → 48.93% - 2-3: PuLID 0.7 + Redux 0.3 → 38.52% (Redux 낮으면 파스텔 감소) - 2-4: PuLID 0.85 + Redux 0.5 → 49.99% ★ R2 best & 전체 best - 2-5: 한옥 spring (cherry blossom) → 44.45% / black 4.3% R1 best (R1-4 48.5%) → R2-4 (49.99%) = +1.5 pp 갱신 R9 legacy (44.5%) → R2-4 (49.99%) = +5.5 pp 누적 학습: - PuLID weight 0.85 + Redux 0.5 multiply 조합이 0.7/0.5 보다 우월 - Redux 0.3 너무 낮음, 0.7 비슷, 0.85+ 미검증 - Stage 3 Refiner 효과 미미 (48.5% → 48.82%, +0.3pp) CUDA error 회복: - 2-3, 2-4, 2-5 첫 시도 CUDA invalid argument - ComfyUI venv 재시작 후 retry → 3/3 모두 success - 매 ~5건 후 재시작 패턴 확인됨 자산 라이브러리: - models-performance: 8 모델 추적 - workflows: winner_auto 4 (R1-4, R2-1, R2-2, R2-4) + partial 5 - prompts/proven: 7 패턴 (R1, R2 합산) R3 계획: - 3-1: PuLID 0.7 + Redux 0.85 (push redux 더 강) - 3-2: PuLID 0.6 + Redux 0.7 (PuLID 약화) - 3-3: PuLID 0.7 + Redux 0.7 @ 1024x1792 (해상도 증가) - 3-4: Qwen-Image fp8 + Lightning LoRA 60갑자 (fp8 ~20G 완료, GGUF 회피) - 3-5: Pony V6 img2img from byeolyi-ref (대안 ID 방법) 검수: D:\Vault\8460s-image-rd\eval\meeting-103-v2.1-clean-grid.html
2026-05-19 23:36:44 +09:00
<div class="label-flags"><span class="flag flag-ok">clean</span></div>
</div>
</div>
<div class="cell status-success">
<img src="../results/meeting-103-v2.1-clean/round2/2-5_thumb.png" data-original="../results/meeting-103-v2.1-clean/round2/2-5.png" alt="2-5">
<div class="label">
<div class="label-title">R2-2-5: B — Hanok spring variant (cherry blossom)</div>
meeting-103 v2.1 clean R3+R4 ✅: PuLID 0.7+Redux 1.0 = pastel 56.41% (전체 best 갱신) R3 (4/5 success): - 3-1: PuLID 0.7 + Redux 0.85 → 50.47% (R2 best 49.99% 갱신) - 3-2: PuLID 0.6 + Redux 0.7 → 49.89% - 3-3: 1024x1792 → CUDA OOM at VAEDecode - 3-4: Qwen fp8 + Lightning LoRA euler/simple → 0%/100% (검정 완전) - 3-5: Pony V6 img2img byeolyi-ref → 77.3% (img2img 0.65 denoise, 원본 색 보존) R4 (3/5 success, 별이 새 best): - 4-1: PuLID 0.7 + Redux 1.0 (max) → 56.41% ★ 전체 best - 4-2: PuLID 0.85 + Redux 0.7 → 50.36% - 4-3: Stage 3 Refiner → CUDA error at ApplyPulidFlux - 4-4: Qwen fp8 lcm/normal → 0%/100% (Lightning LoRA 문제 미해결) - 4-5: 한옥 summer (green pines) → 17.7%/3.97% 학습: - Redux strength 0.5 → 0.7 → 0.85 → 1.0 단조 증가 (56.41% sweet spot) - PuLID 0.7 fixed, Redux 1.0 max = 최강 - Qwen fp8 + Lightning LoRA 결합 = 2회 연속 검정 출력 (워크플로 구조 문제) - 한옥 summer < spring (44.45%) — 색감 자체가 차분 진행: R9 legacy 44.5% → R1 48.5% → R2 49.99% → R3 50.47% → R4 56.41% (+11.91pp) 자산 라이브러리: - workflows: winner_auto 7 + partial 7 - prompts/failed: 9 패턴 (Qwen 계열 2건 추가)
2026-05-19 23:49:55 +09:00
<div class="label-meta">models=flux-dev</div>
<div class="label-auto-eval">pastel=44.45% / black=4.3%</div>
<div class="label-flags"><span class="flag flag-ok">clean</span></div>
</div>
</div>
<div class="cell status-success">
<img src="../results/meeting-103-v2.1-clean/round3/3-1_thumb.png" data-original="../results/meeting-103-v2.1-clean/round3/3-1.png" alt="3-1">
<div class="label">
<div class="label-title">R3-3-1: A — PuLID 0.7 + Redux 0.85</div>
<div class="label-meta">models=flux-dev, pulid-flux, flux-redux</div>
<div class="label-auto-eval">pastel=50.47% / black=0.24%</div>
<div class="label-flags"><span class="flag flag-ok">clean</span></div>
</div>
</div>
<div class="cell status-success">
<img src="../results/meeting-103-v2.1-clean/round3/3-2_thumb.png" data-original="../results/meeting-103-v2.1-clean/round3/3-2.png" alt="3-2">
<div class="label">
<div class="label-title">R3-3-2: A — PuLID 0.6 + Redux 0.7</div>
<div class="label-meta">models=flux-dev, pulid-flux, flux-redux</div>
<div class="label-auto-eval">pastel=49.89% / black=0.22%</div>
<div class="label-flags"><span class="flag flag-ok">clean</span></div>
</div>
</div>
<div class="cell status-error">
<div class="no-image">ERROR — no image</div>
<div class="label">
<div class="label-title">R3-3-3: A — PuLID 0.7 + Redux 0.7 @ 1024x1792</div>
<div class="label-meta">models=flux-dev, pulid-flux, flux-redux</div>
<div class="label-auto-eval">error: VAEDecode</div>
<div class="label-flags"><span class="flag flag-warn">status-error</span></div>
</div>
</div>
<div class="cell status-success">
<img src="../results/meeting-103-v2.1-clean/round3/3-4_thumb.png" data-original="../results/meeting-103-v2.1-clean/round3/3-4.png" alt="3-4">
<div class="label">
<div class="label-title">R3-3-4: C — 60갑자 Qwen-Image fp8 + Lightning LoRA 4-step</div>
<div class="label-meta">models=qwen-image-2.0, lightning-lora-qwen</div>
<div class="label-auto-eval">pastel=0.0% / black=100.0%</div>
<div class="label-flags"><span class="flag flag-warn">black-suspicious</span><span class="flag flag-warn">pastel-low</span></div>
</div>
</div>
<div class="cell status-success">
<img src="../results/meeting-103-v2.1-clean/round3/3-5_thumb.png" data-original="../results/meeting-103-v2.1-clean/round3/3-5.png" alt="3-5">
<div class="label">
<div class="label-title">R3-3-5: A — Pony V6 img2img from byeolyi-ref (denoise 0.65)</div>
<div class="label-meta">models=pony-v6</div>
<div class="label-auto-eval">pastel=77.3% / black=0.03%</div>
<div class="label-flags"><span class="flag flag-ok">clean</span></div>
</div>
</div>
<div class="cell status-success">
<img src="../results/meeting-103-v2.1-clean/round4/4-1_thumb.png" data-original="../results/meeting-103-v2.1-clean/round4/4-1.png" alt="4-1">
<div class="label">
<div class="label-title">R4-4-1: A — PuLID 0.7 + Redux 1.0 (max)</div>
<div class="label-meta">models=flux-dev, pulid-flux, flux-redux</div>
<div class="label-auto-eval">pastel=56.41% / black=0.2%</div>
<div class="label-flags"><span class="flag flag-ok">clean</span></div>
</div>
</div>
<div class="cell status-success">
<img src="../results/meeting-103-v2.1-clean/round4/4-2_thumb.png" data-original="../results/meeting-103-v2.1-clean/round4/4-2.png" alt="4-2">
<div class="label">
<div class="label-title">R4-4-2: A — PuLID 0.85 + Redux 0.7 (combine best)</div>
<div class="label-meta">models=flux-dev, pulid-flux, flux-redux</div>
<div class="label-auto-eval">pastel=50.36% / black=0.86%</div>
<div class="label-flags"><span class="flag flag-ok">clean</span></div>
</div>
</div>
<div class="cell status-error">
<div class="no-image">ERROR — no image</div>
<div class="label">
<div class="label-title">R4-4-3: A — PuLID 0.7 + Redux 0.85 + Stage 3 Refiner</div>
<div class="label-meta">models=flux-dev, pulid-flux, flux-redux, flux-dev-refiner</div>
<div class="label-auto-eval">error: ApplyPulidFlux</div>
<div class="label-flags"><span class="flag flag-warn">status-error</span></div>
</div>
</div>
<div class="cell status-success">
<img src="../results/meeting-103-v2.1-clean/round4/4-4_thumb.png" data-original="../results/meeting-103-v2.1-clean/round4/4-4.png" alt="4-4">
<div class="label">
<div class="label-title">R4-4-4: C — 60갑자 Qwen-Image fp8 + Lightning LoRA (lcm/normal cfg=1)</div>
<div class="label-meta">models=qwen-image-2.0, lightning-lora-qwen</div>
<div class="label-auto-eval">pastel=0.0% / black=100.0%</div>
<div class="label-flags"><span class="flag flag-warn">black-suspicious</span><span class="flag flag-warn">pastel-low</span></div>
</div>
</div>
<div class="cell status-success">
<img src="../results/meeting-103-v2.1-clean/round4/4-5_thumb.png" data-original="../results/meeting-103-v2.1-clean/round4/4-5.png" alt="4-5">
<div class="label">
<div class="label-title">R4-4-5: B — Hanok summer (green pines, blue sky)</div>
<div class="label-meta">models=flux-dev</div>
<div class="label-auto-eval">pastel=17.72% / black=3.97%</div>
meeting-103 v2.1 clean R2 ✅5/5: PuLID 0.85+Redux 0.5 = pastel 49.99% (전체 best 갱신) R2 5조합 (5/5 success, retry 2회 포함): - 2-1: PuLID 0.7 + Redux 0.5 + Stage 3 Refiner (denoise 0.3) → 48.82% (Stage 3 효과 미미) - 2-2: PuLID 0.7 + Redux 0.7 → 48.93% - 2-3: PuLID 0.7 + Redux 0.3 → 38.52% (Redux 낮으면 파스텔 감소) - 2-4: PuLID 0.85 + Redux 0.5 → 49.99% ★ R2 best & 전체 best - 2-5: 한옥 spring (cherry blossom) → 44.45% / black 4.3% R1 best (R1-4 48.5%) → R2-4 (49.99%) = +1.5 pp 갱신 R9 legacy (44.5%) → R2-4 (49.99%) = +5.5 pp 누적 학습: - PuLID weight 0.85 + Redux 0.5 multiply 조합이 0.7/0.5 보다 우월 - Redux 0.3 너무 낮음, 0.7 비슷, 0.85+ 미검증 - Stage 3 Refiner 효과 미미 (48.5% → 48.82%, +0.3pp) CUDA error 회복: - 2-3, 2-4, 2-5 첫 시도 CUDA invalid argument - ComfyUI venv 재시작 후 retry → 3/3 모두 success - 매 ~5건 후 재시작 패턴 확인됨 자산 라이브러리: - models-performance: 8 모델 추적 - workflows: winner_auto 4 (R1-4, R2-1, R2-2, R2-4) + partial 5 - prompts/proven: 7 패턴 (R1, R2 합산) R3 계획: - 3-1: PuLID 0.7 + Redux 0.85 (push redux 더 강) - 3-2: PuLID 0.6 + Redux 0.7 (PuLID 약화) - 3-3: PuLID 0.7 + Redux 0.7 @ 1024x1792 (해상도 증가) - 3-4: Qwen-Image fp8 + Lightning LoRA 60갑자 (fp8 ~20G 완료, GGUF 회피) - 3-5: Pony V6 img2img from byeolyi-ref (대안 ID 방법) 검수: D:\Vault\8460s-image-rd\eval\meeting-103-v2.1-clean-grid.html
2026-05-19 23:36:44 +09:00
<div class="label-flags"><span class="flag flag-ok">clean</span></div>
meeting-103 v2.1 clean R1 ✅5/5: PuLID + Redux = pastel 48.5% (R9 best 갱신) R0 환경 점검: - VRAM 11.7/12.4GB free (94%) — 해소 확인 - pytorch 2.12.0+cu130 + sageattention venv 검증 - 모델 인벤토리: Flux Dev Q5 GGUF, PuLID, Redux, byeolyi-ref-916 확보 - fp8 다운로드 .part 0 bytes (재개 대기) R1 clean 5조합 (5/5 success): - 1-1: PuLID w0.85 mean → pastel 25.7% / black 1.95% - 1-2: PuLID w0.85 mean delayed_start 0.05 → pastel 10.0% / black 14.0% (어두움) - 1-3: PuLID w1.0 mean → pastel 27.5% / black 0.79% - 1-4: PuLID w0.7 + Redux 0.5 → pastel 48.5% / black 0.27% ★ R1 best - 1-5: 한옥 baseline Flux Dev 768² → pastel 12.9% / black 11.7% CUDA error 회복: - 1-2 retry concat → CUDA invalid argument → 후속 1-2~1-5 모두 실패 - ComfyUI venv 재시작 (sage + pytorch 2.12) → 1-2 변형 (delayed_start)로 대체 - 재시도 4/4 모두 success 이전 R9 best (44.5%) → R1-4 (48.5%) = +4 pp 갱신 자산 라이브러리 누적: - models-performance.json: 7 모델 추적 (flux-dev, pulid-flux, flux-redux 등) - workflows: winner_auto 1 + partial 2 - prompts/proven: illustration_auto 패턴 3건 - prompts/failed: pastel-low + CUDA error 패턴 - prompts/negative: A_byeolyi 패턴 etc. 검수: - Vault: D:\Vault\8460s-image-rd\eval\meeting-103-v2.1-clean-grid.html - Gitea: results/meeting-103-v2.1-clean/round1/
2026-05-19 23:17:19 +09:00
</div>
</div>
</div>
</body></html>
Reference in New Issue Copy Permalink