window.BENCH_DATA = {"bench": "object", "bench_cn": "刚性物体 / 产品(鞋、包、道具、硬件)", "bench_en": "Rigid objects / products (shoes, bags, props, hardware)", "cases": [{"id": "PB0", "cn": "基线", "en": "Baseline", "prompt": "{t}, studio product photo, soft even lighting, plain background"}, {"id": "PB1", "cn": "夜景", "en": "Night scene", "prompt": "{t}, on wet asphalt at night, neon reflections, shallow depth of field"}, {"id": "PB2", "cn": "特写", "en": "Close-up", "prompt": "{t}, extreme close-up, 85mm macro, f/1.8"}, {"id": "PB3", "cn": "俯视", "en": "Top-down", "prompt": "{t}, viewed from directly above, on a flat surface"}, {"id": "PB4", "cn": "纯黑底", "en": "Pure black", "prompt": "{t}, on pure black seamless background, hard rim light, extreme minimalism"}, {"id": "PB5", "cn": "油画", "en": "Oil painting", "prompt": "{t}, oil painting, impasto, visible brush strokes"}, {"id": "PB6", "cn": "裸触发", "en": "Trigger only", "prompt": "{t}"}, {"id": "PB7", "cn": "远景", "en": "Wide shot", "prompt": "{t}, full body, side view, wide shot"}], "terms": [{"k": "WW", "cn": "可用强度区间。下界(还认得出)到上界(还听 prompt 话)之间的宽度,越宽越好用。", "en": "Usable strength window: width between the lower bound (still recognisable) and upper bound (prompt still obeyed). Wider is better."}, {"k": "0.00", "cn": "原始模型对照列。不加载 LoRA,用来判断这条 LoRA 到底改了什么。", "en": "Base reference column. No LoRA loaded, used to see what the LoRA actually changes."}, {"k": "PB", "cn": "8 个固定探针,每个只隔离一个变量(场景/尺度/视角/背景/风格/裸词/构图)。", "en": "8 fixed probes, each isolating one variable (scene / scale / angle / background / style / bare trigger / framing)."}, {"k": "CONT", "cn": "污染指纹。和原始模型比,色调偏了多少、锐度变了多少。越小越干净。", "en": "Contamination: how much tone and sharpness drift from the base model. Lower is cleaner."}, {"k": "IDS", "cn": "跨种子一致性。同一格换随机种子,出图还像不像。", "en": "Seed consistency: whether different seeds in the same cell still look alike."}], "report": {"cn": "时间: 2026-11-20 09:12:46\n\n[1] Pipeline\n\n UNET moss_image_sft_int8_convrot.safetensors\n TE qwen3.5_9b_int8_convrot.safetensors (moss_image)\n VAE flux2-vae_small.safetensors\n Steps/CFG 8 / 1 / euler_ancestral / beta\n Size 888 x 1184\n Fixed LoRA 8-steps-dmd2-lora.safetensors @ 1\n Target LoRA DIY\\MOSS\\Heels\\ClearSingleStrapSandals_MI_e4_onetrainer.safetensors\n Trigger css_1, clear single-strap heeled sandals\n Seeds 12345\n\n[2] 自动指标\n\n 强度 色调偏移 锐度比 结论\n 0.40 0.452 0.833 明显改色调,且变得更柔/更平\n 0.60 0.658 0.649 明显改色调,且变得更柔/更平\n 0.75 0.624 0.726 明显改色调,且变得更柔/更平\n 0.85 0.681 0.629 明显改色调,且变得更柔/更平\n 1.00 0.73 0.636 明显改色调,且变得更柔/更平\n\n 色调偏移 = 0 表示和原始模型色调完全一致;0.5 以上说明这条 LoRA 会明显改变整体观感。\n 锐度比 = 1.0 表示细节量没变;1.5 表示高频细节比原始模型多 50%。\n\n[3] 需人工判定\n\n 1) 下界:从 0.40 往上,哪一档开始「一眼认得出是这个东西」 -> ____\n 2) 上界:看 PB4 那一行,哪一档开始失效/被拽回训练分布 -> ____\n 3) WW = 上界 - 下界 -> ____\n 4) PAS@0.75:这些用例的约束有几个成立(共 6 个)\n PB1 / PB2 / PB3 / PB4 / PB5 / PB7\n 5) 分档:WW>=0.15 且 PAS>=0.6 -> A;WW>=0.10 -> B;否则 C\n\n[4] 建议\n\n· 色调偏移偏大(0.73 @ 1.00)。发布时建议写明「会改变整体色调」,或提供低强度用法。\n· 出图比原始模型更柔(0.636x)。如果训练集是高清实拍,说明数据锐度不够。\n· 蒸馏管线(低步数 / CFG≈1):风格迁移和长句服从天然弱,测出来的瓶颈别算在 LoRA 头上。\n· 管线里还有固定 LoRA(8-steps-dmd2-lora.safetensors)。它和 Target 共享强度预算 —— 记录 WW 时务必一起写。\n\n[5] 产出\n 图片目录 : G:/StableDiffusion/ComfyUI-aki-v1.4/output\\bench\\object\\ClearSingleStrapSandals_MI_e4_onetrainer_00014\n 对照图 : G:/StableDiffusion/ComfyUI-aki-v1.4/output\\bench\\object\\ClearSingleStrapSandals_MI_e4_onetrainer_00014\\_sheet_cn.jpg\n 对照图 : G:/StableDiffusion/ComfyUI-aki-v1.4/output\\bench\\object\\ClearSingleStrapSandals_MI_e4_onetrainer_00014\\_sheet_en.jpg\n====================================================================\n", "en": "time: 2026-11-20 09:12:46\n\n[1] Pipeline\n\n UNET moss_image_sft_int8_convrot.safetensors\n TE qwen3.5_9b_int8_convrot.safetensors (moss_image)\n VAE flux2-vae_small.safetensors\n Steps/CFG 8 / 1 / euler_ancestral / beta\n Size 888 x 1184\n Fixed LoRA 8-steps-dmd2-lora.safetensors @ 1\n Target LoRA DIY\\MOSS\\Heels\\ClearSingleStrapSandals_MI_e4_onetrainer.safetensors\n Trigger css_1, clear single-strap heeled sandals\n Seeds 12345\n\n[2] Auto metrics\n\n strength tone sharp reading\n 0.40 0.452 0.833 tone clearly shifted, softer / flatter\n 0.60 0.658 0.649 tone clearly shifted, softer / flatter\n 0.75 0.624 0.726 tone clearly shifted, softer / flatter\n 0.85 0.681 0.629 tone clearly shifted, softer / flatter\n 1.00 0.73 0.636 tone clearly shifted, softer / flatter\n\n tone = 0 means identical tone to base; above 0.5 the LoRA visibly changes the overall look.\n sharp = 1.0 same detail level; 1.5 means 50% more high-frequency detail than base.\n\n[3] Manual judgement\n\n 1) Lower bound: scanning up from 0.40, where is it obviously this thing? -> ____\n 2) Upper bound: on the PB4 row, where does it break or snap back? -> ____\n 3) WW = upper - lower -> ____\n 4) PAS@0.75: how many of these cases keep their constraints (6 total)\n PB1 / PB2 / PB3 / PB4 / PB5 / PB7\n 5) Grade: WW>=0.15 and PAS>=0.6 -> A; WW>=0.10 -> B; else C\n\n[4] Recommendations\n\n· Tone drift is high (0.73 @ 1.00). Say so in the model card, or recommend a lower strength.\n· Softer than base (0.636x) — if the dataset was hi-res photography, it may be under-sharp.\n· Distilled pipeline (low steps / CFG~1): style transfer and long-prompt adherence are inherently weak — do not blame the LoRA.\n· Pipeline also loads fixed LoRA(s): 8-steps-dmd2-lora.safetensors. They share the strength budget — always record them with the WW.\n\n[5] Output\n images : G:/StableDiffusion/ComfyUI-aki-v1.4/output\\bench\\object\\ClearSingleStrapSandals_MI_e4_onetrainer_00014\n sheet : G:/StableDiffusion/ComfyUI-aki-v1.4/output\\bench\\object\\ClearSingleStrapSandals_MI_e4_onetrainer_00014\\_sheet_cn.jpg\n sheet : G:/StableDiffusion/ComfyUI-aki-v1.4/output\\bench\\object\\ClearSingleStrapSandals_MI_e4_onetrainer_00014\\_sheet_en.jpg\n====================================================================\n"}, "runs": [{"tag": "ClearSingleStrapSandals_MI_e4_onetrainer_00014", "bench": "object", "time": "2026-11-20 09:12", "unet": "moss_image_sft_int8_convrot.safetensors", "te": "qwen3.5_9b_int8_convrot.safetensors", "clip_type": "moss_image", "vae": "flux2-vae_small.safetensors", "steps": 8, "cfg": 1, "sampler": "euler_ancestral", "scheduler": "beta", "size": [888, 1184], "extra": [{"name": "8-steps-dmd2-lora.safetensors", "strength": 1}], "target": "DIY\\MOSS\\Heels\\ClearSingleStrapSandals_MI_e4_onetrainer.safetensors", "trigger": "css_1, clear single-strap heeled sandals", "seeds": [12345], "cells": {"PB0": {"0.00": "mock_data/PB0_s000_sd12345_00001_.webp", "0.40": "mock_data/PB0_s040_sd12345_00001_.webp", "0.60": "mock_data/PB0_s060_sd12345_00001_.webp", "0.75": "mock_data/PB0_s075_sd12345_00001_.webp", "0.85": "mock_data/PB0_s085_sd12345_00001_.webp", "1.00": "mock_data/PB0_s100_sd12345_00001_.webp"}, "PB1": {"0.00": "mock_data/PB1_s000_sd12345_00001_.webp", "0.40": "mock_data/PB1_s040_sd12345_00001_.webp", "0.60": "mock_data/PB1_s060_sd12345_00001_.webp", "0.75": "mock_data/PB1_s075_sd12345_00001_.webp", "0.85": "mock_data/PB1_s085_sd12345_00001_.webp", "1.00": "mock_data/PB1_s100_sd12345_00001_.webp"}, "PB2": {"0.00": "mock_data/PB2_s000_sd12345_00001_.webp", "0.40": "mock_data/PB2_s040_sd12345_00001_.webp", "0.60": "mock_data/PB2_s060_sd12345_00001_.webp", "0.75": "mock_data/PB2_s075_sd12345_00001_.webp", "0.85": "mock_data/PB2_s085_sd12345_00001_.webp", "1.00": "mock_data/PB2_s100_sd12345_00001_.webp"}, "PB3": {"0.00": "mock_data/PB3_s000_sd12345_00001_.webp", "0.40": "mock_data/PB3_s040_sd12345_00001_.webp", "0.60": "mock_data/PB3_s060_sd12345_00001_.webp", "0.75": "mock_data/PB3_s075_sd12345_00001_.webp", "0.85": "mock_data/PB3_s085_sd12345_00001_.webp", "1.00": "mock_data/PB3_s100_sd12345_00001_.webp"}, "PB4": {"0.00": "mock_data/PB4_s000_sd12345_00001_.webp", "0.40": "mock_data/PB4_s040_sd12345_00001_.webp", "0.60": "mock_data/PB4_s060_sd12345_00001_.webp", "0.75": "mock_data/PB4_s075_sd12345_00001_.webp", "0.85": "mock_data/PB4_s085_sd12345_00001_.webp", "1.00": "mock_data/PB4_s100_sd12345_00001_.webp"}, "PB5": {"0.00": "mock_data/PB5_s000_sd12345_00001_.webp", "0.40": "mock_data/PB5_s040_sd12345_00001_.webp", "0.60": "mock_data/PB5_s060_sd12345_00001_.webp", "0.75": "mock_data/PB5_s075_sd12345_00001_.webp", "0.85": "mock_data/PB5_s085_sd12345_00001_.webp", "1.00": "mock_data/PB5_s100_sd12345_00001_.webp"}, "PB6": {"0.00": "mock_data/PB6_s000_sd12345_00001_.webp", "0.40": "mock_data/PB6_s040_sd12345_00001_.webp", "0.60": "mock_data/PB6_s060_sd12345_00001_.webp", "0.75": "mock_data/PB6_s075_sd12345_00001_.webp", "0.85": "mock_data/PB6_s085_sd12345_00001_.webp", "1.00": "mock_data/PB6_s100_sd12345_00001_.webp"}, "PB7": {"0.00": "mock_data/PB7_s000_sd12345_00001_.webp", "0.40": "mock_data/PB7_s040_sd12345_00001_.webp", "0.60": "mock_data/PB7_s060_sd12345_00001_.webp", "0.75": "mock_data/PB7_s075_sd12345_00001_.webp", "0.85": "mock_data/PB7_s085_sd12345_00001_.webp", "1.00": "mock_data/PB7_s100_sd12345_00001_.webp"}}}]};