Datasets:
image imagewidth (px) 4.1k 4.1k | id stringlengths 21 21 | question stringclasses 8
values | title stringclasses 8
values | mode stringclasses 2
values | seed int64 18.7k 90.8k | recipe stringclasses 2
values | prompt stringclasses 8
values | hard_checks listlengths 16 16 | hard_score float64 57.5 80 | coherence_score float64 5 10 | aesthetic_score float64 5 10 | total_score float64 70 100 | strict_pass bool 2
classes | notes_zh stringlengths 48 125 | backend_seconds float64 86.9 372 | reference_files listlengths 0 2 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Q1-18731-base-sixstep | Q1 | 数量、属性绑定与空间遮挡 | t2i | 18,731 | base-sixstep | A precise, photorealistic tabletop product photograph, square frame, plain warm-gray background and neutral soft studio lighting. Exactly seven small painted wooden objects sit on a single rectangular cream presentation mat. There are no people, letters, mirrors or additional objects. The viewpoint is slightly above th... | [
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5
] | 80 | 10 | 7.5 | 97.5 | true | 7物体和绑定均成立,黑球在后右,黄锥遮挡可见;粗糙涂木质感自然。两图构图均清楚但偏常规商品陈列。 | 178.013 | [] | |
Q1-18731-pdmd-twostep | Q1 | 数量、属性绑定与空间遮挡 | t2i | 18,731 | pdmd-twostep | A precise, photorealistic tabletop product photograph, square frame, plain warm-gray background and neutral soft studio lighting. Exactly seven small painted wooden objects sit on a single rectangular cream presentation mat. There are no people, letters, mirrors or additional objects. The viewpoint is slightly above th... | [
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5
] | 80 | 7.5 | 7.5 | 95 | true | 7物体和绑定均成立。黑球在后右,按物体位置而非轮廓最右端判通过;部分表面带规则织纹、偏亮,木材可信度稍弱。 | 86.918 | [] | |
Q2-18731-base-sixstep | Q2 | 精确中文、拉丁字符与排版 | t2i | 18,731 | base-sixstep | A beautifully restrained fictional bookshop announcement printed on one perfectly front-facing square sheet of warm-white paper. The sheet fills the frame without perspective distortion. This is a finished typographic design, not a photograph of a busy shop. Use sharp charcoal typography, a small muted-teal rule and ge... | [
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5
] | 80 | 10 | 10 | 100 | true | 中文、标点、时间、A0B1-2026及O0 I1 lL均清楚。零比O窄,I/1/l/L具可见字形差异;五块信息完整,页脚对齐,居中层级克制。 | 177.314 | [] | |
Q2-18731-pdmd-twostep | Q2 | 精确中文、拉丁字符与排版 | t2i | 18,731 | pdmd-twostep | A beautifully restrained fictional bookshop announcement printed on one perfectly front-facing square sheet of warm-white paper. The sheet fills the frame without perspective distortion. This is a finished typographic design, not a photograph of a busy shop. Use sharp charcoal typography, a small muted-teal rule and ge... | [
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5
] | 80 | 10 | 7.5 | 97.5 | true | 字面文字、易混淆字符、层级与页脚均成立;排版较常规,时间块偏左且多一条装饰线,作为审美差异而非临时新增硬扣分。 | 87.041 | [] | |
Q3-18731-base-sixstep | Q3 | 图表的数据真实性 | t2i | 18,731 | base-sixstep | A flat, front-facing square statistical information graphic for a fictional classroom experiment. The title is '样本计数'. Crisp vector-style lines on a white background, dark ink and three muted teal bars; no texture, perspective, people, shadows or decorative chart elements. This is actual quantitative information, not a... | [
5,
2.5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5
] | 77.5 | 7.5 | 7.5 | 92.5 | false | 图的刻度、网格、三柱高度及2:5:8比例正确。表格却在左右两列都重复A|2、B|5、C|8,未实现组别列和数量列的分工;数值本身仍一致,不重复扣成额外虚构数据。 | 178.646 | [] | |
Q3-18731-pdmd-twostep | Q3 | 图表的数据真实性 | t2i | 18,731 | pdmd-twostep | A flat, front-facing square statistical information graphic for a fictional classroom experiment. The title is '样本计数'. Crisp vector-style lines on a white background, dark ink and three muted teal bars; no texture, perspective, people, shadows or decorative chart elements. This is actual quantitative information, not a... | [
5,
5,
5,
5,
5,
0,
5,
0,
2.5,
2.5,
5,
0,
2.5,
5,
5,
5
] | 57.5 | 5 | 7.5 | 70 | false | 数值、表格、合计和均值正确;但数轴不等距,水平网格比整数刻度多且不对应。A/B柱未准确落在2/5,C到8;柱高比例偏差见像素诊断。 | 89.363 | [] | |
Q3-90817-base-sixstep | Q3 | 图表的数据真实性 | t2i | 90,817 | base-sixstep | A flat, front-facing square statistical information graphic for a fictional classroom experiment. The title is '样本计数'. Crisp vector-style lines on a white background, dark ink and three muted teal bars; no texture, perspective, people, shadows or decorative chart elements. This is actual quantitative information, not a... | [
5,
2.5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5
] | 77.5 | 7.5 | 7.5 | 92.5 | false | 柱高、等距刻度、网格与2:5:8比例再次正确;但两列表格再次左右重复A|2、B|5、C|8,没有实现列语义分离。图表稳定性不等于整题完整合格。 | 189.82 | [] | |
Q3-90817-pdmd-twostep | Q3 | 图表的数据真实性 | t2i | 90,817 | pdmd-twostep | A flat, front-facing square statistical information graphic for a fictional classroom experiment. The title is '样本计数'. Crisp vector-style lines on a white background, dark ink and three muted teal bars; no texture, perspective, people, shadows or decorative chart elements. This is actual quantitative information, not a... | [
5,
2.5,
5,
5,
5,
2.5,
0,
5,
5,
5,
5,
2.5,
5,
5,
5,
5
] | 67.5 | 5 | 5 | 77.5 | false | 文字数值正确,柱顶较接近2/5/8网格,但彩色柱底抬离0基线,形成台阶/白色底座;整数间距仍不完全均匀,柱高比例有偏差。表格有多余竖线、横向断续伪影,A/2等仍可辨但不能当成干净的两列表。 | 91.781 | [] | |
Q4-18731-base-sixstep | Q4 | 光学、镜像与接触物理 | t2i | 18,731 | base-sixstep | A carefully arranged real optics demonstration photographed at table height in a square frame. It contains one and only one physical cube, one clear cylindrical glass of water, one straight striped straw and one flat white arrow painted on the dark tabletop. No people or lettering.
A single upright rectangular plane mi... | [
5,
5,
5,
5,
5,
5,
2.5,
0,
5,
5,
2.5,
2.5,
5,
5,
5,
5
] | 67.5 | 7.5 | 7.5 | 82.5 | false | 直接红脸/镜中蓝脸成立,但桌箭头向右、镜箭头却向左,违反此镜面几何。杯子同样被镜中显示。水面有错位,不能仅凭看起来弯了判为正确折射;实际直管与角度仅部分确认。 | 184.306 | [] | |
Q4-18731-pdmd-twostep | Q4 | 光学、镜像与接触物理 | t2i | 18,731 | pdmd-twostep | A carefully arranged real optics demonstration photographed at table height in a square frame. It contains one and only one physical cube, one clear cylindrical glass of water, one straight striped straw and one flat white arrow painted on the dark tabletop. No people or lettering.
A single upright rectangular plane mi... | [
5,
5,
5,
5,
5,
5,
2.5,
5,
5,
5,
2.5,
2.5,
5,
5,
5,
5
] | 72.5 | 7.5 | 7.5 | 87.5 | false | 红色直接前脸、蓝色镜中背脸及向右箭头均成立。杯子却进入了本应排除的镜中区域;反射物体比例也不够可信。水面有可辨错位,但水上曲线到水下直线的处理无法确认是正确直管折射,只给角度与非真实弯曲部分分。 | 90.195 | [] | |
Q4-90817-base-sixstep | Q4 | 光学、镜像与接触物理 | t2i | 90,817 | base-sixstep | A carefully arranged real optics demonstration photographed at table height in a square frame. It contains one and only one physical cube, one clear cylindrical glass of water, one straight striped straw and one flat white arrow painted on the dark tabletop. No people or lettering.
A single upright rectangular plane mi... | [
5,
5,
5,
5,
5,
5,
2.5,
5,
5,
5,
2.5,
2.5,
5,
5,
5,
5
] | 72.5 | 5 | 7.5 | 85 | false | 红前脸、蓝色反射脸与两个右向箭头成立。镜子左边仍露出杯的反射;反射中的吸管条纹颜色与直接场景不同,且方块顶面额外为蓝,降低连贯性。水面有明显转折与局部偏移,但水下角度和是否真实弯管仍无法确认。光从左上、右向阴影、实体接地基本合理。 | 189.309 | [] | |
Q4-90817-pdmd-twostep | Q4 | 光学、镜像与接触物理 | t2i | 90,817 | pdmd-twostep | A carefully arranged real optics demonstration photographed at table height in a square frame. It contains one and only one physical cube, one clear cylindrical glass of water, one straight striped straw and one flat white arrow painted on the dark tabletop. No people or lettering.
A single upright rectangular plane mi... | [
5,
5,
5,
5,
5,
5,
2.5,
5,
5,
5,
2.5,
2.5,
5,
5,
5,
5
] | 72.5 | 7.5 | 7.5 | 87.5 | false | 红色直接前脸、蓝色镜中背脸,直接和反射箭头均向右;杯子与吸管却完整进入本应排除的镜场,反射杯的尺度也不够可信。水面可见弯曲/偏移,水上水下轴线角度变化不明确,无法确认符合直管折射,相关两项仅给部分分。光影与接地基本成立;方块上面额外变蓝是连贯性缺点。 | 90.65 | [] | |
Q5-18731-base-sixstep | Q5 | 四只手的协作与归属 | t2i | 18,731 | base-sixstep | An intimate naturalistic editorial photograph, square composition, of exactly two adults wrapping one small gift at a wooden table. On the viewer's left is a woman in a red cardigan; on the right is a man in a plain gray shirt. No other person, mirror, lettering or decorative clutter appears. The camera is at a slightl... | [
5,
5,
5,
2.5,
5,
2.5,
0,
5,
5,
5,
2.5,
5,
2.5,
5,
5,
5
] | 65 | 7.5 | 5 | 77.5 | false | 四手、单盒及红/灰衣服成立,承托与悬空阴影基本合理。但两人头部被完全裁掉,无法核验共同凝视;一只底部支撑手的灰袖/腕归属遮挡。系结处已有部分成形蝴蝶结,左捏环/右穿自由端的具体步骤不清,只给部分分。 | 181.695 | [] | |
Q5-18731-pdmd-twostep | Q5 | 四只手的协作与归属 | t2i | 18,731 | pdmd-twostep | An intimate naturalistic editorial photograph, square composition, of exactly two adults wrapping one small gift at a wooden table. On the viewer's left is a woman in a red cardigan; on the right is a man in a plain gray shirt. No other person, mirror, lettering or decorative clutter appears. The camera is at a slightl... | [
5,
5,
5,
2.5,
5,
2.5,
2.5,
5,
5,
5,
2.5,
5,
2.5,
5,
5,
5
] | 67.5 | 7.5 | 7.5 | 82.5 | false | 四手与单盒成立,女子看向礼盒,但男子眼睛出框,双人凝视只能部分确认。女生抓两条向上的带子,盒面已有完整双环,未清楚表现指定的左环右穿动作;底部另一支撑腕也被挡住。 | 92.184 | [] | |
Q6-18731-base-sixstep | Q6 | 平面与三维的同一空间 | t2i | 18,731 | base-sixstep | A single square architectural presentation divided into exactly two equal panels, one above the other, on a clean white page. The upper panel is titled 'PLAN' and is a true orthographic top-down floor plan. The lower panel is titled 'VIEW' and is an open-roof isometric rendering of the exact same small rectangular room... | [
5,
5,
2.5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5
] | 77.5 | 10 | 7.5 | 95 | false | 两图共享1圆桌、3蓝椅、NE植物、北窗、西墙近SW门洞,位置一致,文字克制。下图同样更像近正面开顶透视,未达到明确的等角轴测要求,给部分分。 | 189.822 | [] | |
Q6-18731-pdmd-twostep | Q6 | 平面与三维的同一空间 | t2i | 18,731 | pdmd-twostep | A single square architectural presentation divided into exactly two equal panels, one above the other, on a clean white page. The upper panel is titled 'PLAN' and is a true orthographic top-down floor plan. The lower panel is titled 'VIEW' and is an open-roof isometric rendering of the exact same small rectangular room... | [
5,
5,
2.5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
0
] | 72.5 | 7.5 | 5 | 85 | false | 两图桌椅、植物、北窗和西侧偏南门洞大体一致。下图为近正面俯视的开顶透视,不能完整确认等角轴测。出现大量额外N/W/E及箭头、重影式注记,违反只要PLAN/VIEW/N,也损害成稿感。 | 92.061 | [] | |
Q7-18731-base-sixstep | Q7 | 双参考局部修改与非目标保留 | edit | 18,731 | base-sixstep | The red cube in the foreground of <Picture 1> is <Subject 1>. The uniform body color filling <Picture 2> is <Subject 2>. The three researchers, their faces, hair, clothing, poses and hands, together with their key and the water glass in <Picture 1>, are <Subject 3>. The green sphere, blue cylinder, yellow cone, magenta... | [
0,
0,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5
] | 70 | 10 | 7.5 | 87.5 | false | 方块仍然红色,编辑主目标失败;人物、动作、其他几何、文字和镜中既有错误等基本保留。冻结量表的保留项很多,故总分仍高,不能以总分宣称编辑成功。 | 371.565 | [
"../references/source-lab.png",
"../references/teal.png"
] | |
Q7-18731-pdmd-twostep | Q7 | 双参考局部修改与非目标保留 | edit | 18,731 | pdmd-twostep | The red cube in the foreground of <Picture 1> is <Subject 1>. The uniform body color filling <Picture 2> is <Subject 2>. The three researchers, their faces, hair, clothing, poses and hands, together with their key and the water glass in <Picture 1>, are <Subject 3>. The green sphere, blue cylinder, yellow cone, magenta... | [
2.5,
5,
5,
5,
2.5,
5,
5,
5,
2.5,
5,
5,
5,
5,
5,
2.5,
5
] | 70 | 5 | 5 | 80 | false | 方块改成偏蓝的青色,接近供体但不完全匹配。人物脸部出现明显皱纹/斑点及锐化,绿球、衣物、墙面和垫子材质重绘,整体对比/光照也变;主要几何、文字、镜中内容和构图大体保住。 | 206.015 | [
"../references/source-lab.png",
"../references/teal.png"
] | |
Q8-18731-base-sixstep | Q8 | 叙事、凝视与克制的真实感 | t2i | 18,731 | base-sixstep | A quiet, emotionally specific documentary photograph inside a tiny watch-repair shop on a rainy evening, square frame. Exactly two people occupy one coherent scene. On the left, an elderly woman watchmaker in a muted teal knitted cardigan sits at her worn wooden workbench. On the right, a young adult man in a damp brow... | [
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5
] | 80 | 10 | 7.5 | 97.5 | true | 硬检查同样成立:老妇抬眼、青年垂眼,表仍在老妇掌中,青年双手未接,工具数量与混光对应。湿外套、皱纹及金属细节较浓重,主观上比另一张少一点柔和克制;不据细节密度直接判假。 | 187.94 | [] | |
Q8-18731-pdmd-twostep | Q8 | 叙事、凝视与克制的真实感 | t2i | 18,731 | pdmd-twostep | A quiet, emotionally specific documentary photograph inside a tiny watch-repair shop on a rainy evening, square frame. Exactly two people occupy one coherent scene. On the left, an elderly woman watchmaker in a muted teal knitted cardigan sits at her worn wooden workbench. On the right, a young adult man in a damp brow... | [
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5,
5
] | 80 | 10 | 10 | 100 | true | 两人、服装、单只打开怀表与3件工具成立。老妇看青年,青年低头朝掌中表,双手仍收拢未接表;老妇轻触表边,克制微笑与停顿可读。左暖灯/右后冷窗、皮肤及毛料木金属真实,背景克制。 | 94.351 | [] |
H3-Image Exam — Eight Tests, Twenty First-Try Images
阅读边界/Scope of claims
这是一组自拟题目的测试数据,仅对本卷题目、指定配方、种子和已生成样本的表现负责,不作为该 LoRA 本身能力优劣的最终评判,也不是模型或 LoRA 排行榜。 分数只描述这些样本对预先冻结检查项的满足程度;不能推出它在其他题目、提示、精度、设备或用途上全面更好/更差。失败也不能单独归因于 LoRA,底模、量化、采样和提示都可能参与影响。
This is a small, task-specific test dataset, not a final judgment of the LoRA's overall quality. Conclusions apply only to the listed tasks, recipes, seeds and generated samples. Scores measure compliance with this frozen rubric; they are neither a general ranking nor an isolated causal estimate of the LoRA's effect. The executor and visual grader are the same AI agent; the assessment is not independent or double-blind.
20 generated images, only 6 fully satisfy every frozen hard check. This small, self-designed diagnostic compares one six-step base recipe with one pDMD two-step recipe on eight tasks. It shares successes and failures, not a general model ranking.
中文详评 · Frozen questions and rubric · Method · Scores · Offline gallery · Reproduction
考试题目、考察范围与判定目标 / Test specification
本卷覆盖八类可见图像能力,每科有预先冻结的 16项硬检查。它没有覆盖所有图像生成/编辑用途,也不进行最佳提示搜索或完整精度上限测试。以下“达标目标”是出题时的要求,不表示输出已经做到。
| 科目与原题 | 要求画什么 | 考察范围 | 本题具体达标目标 |
|---|---|---|---|
| Q1 数量、属性绑定与空间遮挡 | 摆放恰好七个涂漆木制几何体:前排红方块、绿球、蓝圆柱、洋红圆环;后排黄锥、白方块、黑球,按指定顺序与遮挡摆放。 | 精确计数、颜色与形状绑定、前后/左右关系、遮挡、接地与禁止增物。 | 数量和各属性一一对应;黄锥在蓝柱后被部分遮挡;黑球最右;不因遮挡发生融合、漂浮或新增形体。 |
| Q2 精确中文、拉丁字符与排版 | 设计雨夜书店版面,逐字呈现“雨夜书店”“借一本书,等一场晴。”、“OPEN 19:30–23:00”、A0B1-2026及O0 I1 lL。 | 中文及标点、营业时间、连接号、易混字符与大小写、五个文本块的层级和页脚布局。 | 文字无增删替换;O/0、I/1、l/L可辨;时间正确;五块文字位置完整且无额外文字。 |
| Q3 图表的数据真实性 | 制作A=2、B=5、C=8的柱状图与两列表格,显示0至8整数网格、合计15、均值5。 | 定量图形真实性、数轴等距、共同零基线、真实柱高、表格列语义、跨区域数据一致。 | 柱体几何确实表示2:5:8,不能只在错误柱顶写对数字;网格对应刻度;表格正确分离组别与数量。 |
| Q4 光学、镜像与接触物理 | 拍摄一个方块、一杯水、一根直条纹吸管和桌面箭头:直接看红前脸,镜中看蓝背脸;杯子位于镜场之外。 | 平面镜几何、反射与实体区分、裁切镜场、直管水面折射、单一主光与接触阴影。 | 两个箭头都向画面右方;镜中不新增杯子;水面偏移/角度变化可辨且不是实际弯管;物体接地与光影一致。 |
| Q5 四只手的协作与归属 | 两位成人合作系一个礼盒:灰袖男性双手托盒,红袖女性左手捏环、右手拉自由端穿环,两人看结。 | 四只手的所有权、左右手分工、明确动作阶段、丝带连续拓扑、解剖、支撑与凝视。 | 四手能归属到正确袖口;画出指定“正在穿环”的时刻;丝带连续并有合理受力,不能仅用已完成的蝴蝶结替代动作。 |
| Q6 平面与三维的同一空间 | 上下绘制同一房间的PLAN正投影与VIEW开顶轴测:一圆桌、东西南三把蓝椅、东北植物、北墙中窗、西墙近西南门。 | 两视图对象数量与方位、门窗位置、真实投影类型、北向对应和文字排除。 | 两图为同一空间,北侧无椅;家具及开口位置一致;下图应为轴测/等角图;只出现PLAN、VIEW、N。 |
| Q7 双参考局部修改与非目标保留 | 使用两个参考:保留实验室底图,只把前景红方块的体色改为第二张纯色图的青绿色,保留其余内容及底图已有错误。 | 参考角色区分、供体色转移、局部性、人物身份与互动保留、非目标几何/文字/背景保留。 | 核心颜色编辑必须完成,其他区域应保留;不能把供体画布导入,或改脸/物/字。保留项高分不等于主编辑成功。 |
| Q8 叙事、凝视与克制的真实感 | 拍摄雨夜修表店的交付前停顿:左侧老妇穿青绿针织、右侧青年穿棕色湿外套;老妇看青年,青年看她掌中的唯一银怀表。 | 叙事时刻、不同凝视、克制情绪、手与怀表关系、三个工具计数、混合照明与材质真实感。 | 青年尚未接表;老妇看脸、青年看表;怀表及螺丝刀/镊子/米色布数量正确;左暖右冷光和材质服务画面焦点。 |
完整原题与逐项检查 / Full prompts and checks
下列英文提示与运行时完全一致;可见中文保持原字面。每项仅取0/2.5/5(失败或不可见/部分满足或不确定/清楚满足)。每科硬分80,另计连贯10与审美10,只有全部硬检查满分才算完整过关。八科等权前先平均题内种子,图表和光学复测不会增加科目权重。
Q1 数量、属性绑定与空间遮挡 — 完整提示、四组目标与16项检查
原始提示 / Exact prompt
A precise, photorealistic tabletop product photograph, square frame, plain warm-gray background and neutral soft studio lighting. Exactly seven small painted wooden objects sit on a single rectangular cream presentation mat. There are no people, letters, mirrors or additional objects. The viewpoint is slightly above the table, so all seven can be counted and their contact with the mat is visible.
The four objects in the front row, from the viewer's left to right, are one red cube, one green sphere, one blue upright cylinder, and one magenta torus lying flat with its central hole clearly visible. The three objects in the back row, from left to right, are one yellow cone, one white cube, and one black sphere. The yellow cone stands directly behind the blue cylinder, partly occluded by it but with its tip and one sloping side visible. The white cube stands behind and to the right of that cone. The black sphere is farther to the right than every other object. The two cubes have different colors and the two spheres have different colors; no colors migrate between shapes. A small gap separates the red cube from the green sphere. All objects are rigid, correctly shaped and grounded by natural contact shadows. The surface is matte with subtle wood grain, without glossy plastic or decorative clutter.
考察范围: 精确计数、颜色与形状绑定、前后/左右关系、遮挡、接地与禁止增物。
达标目标: 数量和各属性一一对应;黄锥在蓝柱后被部分遮挡;黑球最右;不因遮挡发生融合、漂浮或新增形体。
计数与禁增(20分)
- 直接物体恰好7个
- 前排4个/后排3个可分辨
- 仅2方块、2球、1柱、1锥、1环
- 无额外道具/文字/人物
颜色绑定(20分)
- 红方和白方正确
- 绿球和黑球正确
- 蓝柱和黄锥正确
- 洋红环及中孔正确
空间关系(20分)
- 前排红方→绿球→蓝柱→洋红环
- 黄锥在蓝柱后且部分遮挡
- 白方在黄锥后右方
- 黑球比其他物体更靠右
几何与接触(20分)
- 红方与绿球有间隙
- 圆柱直立且圆环平放
- 7物体均与垫子接地
- 遮挡不造成融合/漂浮/额外形体
Q2 精确中文、拉丁字符与排版 — 完整提示、四组目标与16项检查
原始提示 / Exact prompt
A beautifully restrained fictional bookshop announcement printed on one perfectly front-facing square sheet of warm-white paper. The sheet fills the frame without perspective distortion. This is a finished typographic design, not a photograph of a busy shop. Use sharp charcoal typography, a small muted-teal rule and generous intentional negative space. There are no illustrations or logos.
The exact title at the top is '雨夜书店'. Under it, centered, is the exact sentence '借一本书,等一场晴。'. The next line, smaller but fully readable, is 'OPEN 19:30–23:00'. At the lower left is the exact identifier 'A0B1-2026'. At the lower right is the exact mixed-case character test 'O0 I1 lL'. Preserve capitalization, digit zero versus capital O, digit one versus capital I, lower-case l versus upper-case L, punctuation and the en dash in the opening hours. These five text blocks are the complete visible copy. Their wording is never translated, shortened, duplicated or replaced. The four Chinese title characters dominate; the sentence is second in hierarchy, opening hours third, with the two identifiers aligned to the same footer baseline. Every character fits entirely inside the sheet and remains unobstructed.
考察范围: 中文及标点、营业时间、连接号、易混字符与大小写、五个文本块的层级和页脚布局。
达标目标: 文字无增删替换;O/0、I/1、l/L可辨;时间正确;五块文字位置完整且无额外文字。
中文逐字(20分)
- 标题雨夜二字准确
- 标题书店二字准确
- 借一本书与逗号准确
- 等一场晴及句号准确
英文与时间(20分)
- OPEN大写正确
- 19:30准确
- 23:00准确
- 中间为连接号且整行无增删
易混淆字符(20分)
- A0B1-2026准确
- O0准确且可分辨
- I1准确且可分辨
- lL大小写准确且可分辨
布局与排除(20分)
- 5个文本块无遗漏或重复
- 标题/正文/时间层级成立
- 两个页脚分居左右且对齐
- 全部可读且没有额外标志/文字
Q3 图表的数据真实性 — 完整提示、四组目标与16项检查
原始提示 / Exact prompt
A flat, front-facing square statistical information graphic for a fictional classroom experiment. The title is '样本计数'. Crisp vector-style lines on a white background, dark ink and three muted teal bars; no texture, perspective, people, shadows or decorative chart elements. This is actual quantitative information, not a chart-shaped ornament.
The large upper chart is a vertical bar chart with categories 'A', 'B', 'C' in that left-to-right order. The values are exactly A=2, B=5, C=8. All bars have equal width and the same zero baseline. The vertical axis runs from 0 to 8 with all integer ticks 0,1,2,3,4,5,6,7,8 at equally spaced positions. Thin horizontal gridlines align to every integer tick. The A bar ends exactly at the 2 gridline; the B bar ends exactly at the 5 gridline; the C bar ends exactly at the 8 gridline. The exact number '2', '5', or '8' is printed above its corresponding bar. Category labels sit below the baseline, not inside the bars.
Below the chart is a compact two-column table with headers '组别' and '数量' and exactly three rows 'A | 2', 'B | 5', 'C | 8'. At the bottom are two compact summary fields: '合计 15' and '均值 5'. These are the complete data. Do not add percentages, trends, error bars, extra rows or any invented observations. Every table entry, number, mark and tick must agree with this same three-value dataset.
考察范围: 定量图形真实性、数轴等距、共同零基线、真实柱高、表格列语义、跨区域数据一致。
达标目标: 柱体几何确实表示2:5:8,不能只在错误柱顶写对数字;网格对应刻度;表格正确分离组别与数量。
数据文字(20分)
- A/B/C顺序与顶部2/5/8对应
- 表头组别/数量及三行正确
- 合计15正确
- 均值5正确
数轴(20分)
- 从0到8全部9个刻度
- 刻度等距、方向向上增加
- 0为三柱共同基线
- 网格线逐条对应整数刻度
真实柱高(20分)
- A柱顶到2而非仅标2
- B柱顶到5而非仅标5
- C柱顶到8而非仅标8
- 三柱等宽且柱高比例2:5:8
跨区域一致(20分)
- 图与表的3个数值一致
- 分类标签均在基线下
- 无额外柱/行/百分比/数据
- 所有图表与摘要完整无遮挡
Q4 光学、镜像与接触物理 — 完整提示、四组目标与16项检查
原始提示 / Exact prompt
A carefully arranged real optics demonstration photographed at table height in a square frame. It contains one and only one physical cube, one clear cylindrical glass of water, one straight striped straw and one flat white arrow painted on the dark tabletop. No people or lettering.
A single upright rectangular plane mirror stands at the back of the table, behind the cube. The cube's front face, facing the camera, is red; its rear face, facing the mirror, is blue; its other faces are neutral gray. The camera directly sees the red front face. In the mirror, it sees the blue rear face of that same cube, at the appropriate reflected location and size. The mirror is not a second shelf holding a second physical cube. The glass is placed farther left, outside the mirror's cropped reflective field, so there is no second reflected glass. The white arrow on the horizontal tabletop points toward the viewer's right. Its reflection in the rear mirror also points toward the viewer's right; a rear vertical plane mirror reverses depth here, not the horizontal direction of a tabletop arrow parallel to the mirror.
The glass is partly filled with water with a level horizontal waterline. One physically straight straw passes diagonally through the water surface. Its submerged appearance is clearly offset and changes apparent angle at the waterline because of refraction; it is not an actually bent or broken straw. The red cube and glass rest solidly on the table. A single broad soft light from the viewer's upper left gives consistent highlights and soft cast shadows extending toward the right. The cube's reflected illumination, table contact shadows and transparent glass agree with this one coherent scene. Keep all interfaces and the mirror boundaries clearly visible.
考察范围: 平面镜几何、反射与实体区分、裁切镜场、直管水面折射、单一主光与接触阴影。
达标目标: 两个箭头都向画面右方;镜中不新增杯子;水面偏移/角度变化可辨且不是实际弯管;物体接地与光影一致。
物理实体(20分)
- 直接画面只有1个实体方块
- 只有1杯水和1根吸管
- 镜中方块属于反射而非额外实物
- 无额外人物/文字/道具
镜像几何(20分)
- 直接看见红色前脸
- 镜中看见同方块蓝色背脸
- 位置大小与镜面边界合理
- 桌面箭头及其反射均向画面右方
折射与水(20分)
- 水面水平且与杯壁相交合理
- 吸管在水面出现可辨偏移
- 水下表观角度变化符合折射表现
- 并非真实断管/弯管或两根吸管
光照与接触(20分)
- 主光来自画面左上
- 投影主要向右
- 杯与方块均接地无穿插
- 镜中光照与直接场景一致
Q5 四只手的协作与归属 — 完整提示、四组目标与16项检查
原始提示 / Exact prompt
An intimate naturalistic editorial photograph, square composition, of exactly two adults wrapping one small gift at a wooden table. On the viewer's left is a woman in a red cardigan; on the right is a man in a plain gray shirt. No other person, mirror, lettering or decorative clutter appears. The camera is at a slightly elevated three-quarter angle that clearly shows the four wrists, the working fingers and the gift.
One cream rectangular box is suspended just above the table, held securely by the man's two hands cupping the lower left and lower right corners from underneath. Both of his palms support the bottom of that single box; he does not tie the ribbon. The woman uses her two hands to tie one red ribbon around the same box. Her anatomical left hand pinches a partially formed ribbon loop above the lid, while her anatomical right hand pulls the free ribbon end through that loop. Her forearms remain visibly connected to her own red sleeves. The man's wrists remain connected to his gray sleeves. The ribbon is a single continuous flexible strip that wraps around the box and crosses the lid; the partial bow has not become an unrelated extra object. This is one frozen moment while the knot is tightened, not a sequence.
All four hands have plausible bone structure and finger lengths. Fingers bend around the ribbon or box at actual contact points without fusing into the material. The small gap beneath the box and its shadow on the tabletop make the support arrangement readable. Both people look at the knot, with concentrated but relaxed expressions. Soft window light, ordinary skin texture and restrained fabric detail make the scene convincing rather than an airbrushed advertisement.
考察范围: 四只手的所有权、左右手分工、明确动作阶段、丝带连续拓扑、解剖、支撑与凝视。
达标目标: 四手能归属到正确袖口;画出指定“正在穿环”的时刻;丝带连续并有合理受力,不能仅用已完成的蝴蝶结替代动作。
人数与归属(20分)
- 恰好两人且左右服装正确
- 恰好4只手无额外肢体
- 女两手均连红袖
- 男两手均连灰袖
动作分工(20分)
- 男两手从下方承托同一盒
- 女左手捏环且右手拉自由端
- 两人均看正在打的结
- 没有互换角色或多只手同时占同一位置
拓扑接触(20分)
- 仅1个奶油色盒子
- 红丝带连续绕盒并在盖上交叉
- 环与自由端构成正在系结的一条带
- 接触处无穿模/融合/悬空受力
解剖与支撑(20分)
- 4腕及其所有权清楚
- 可见手指无明显增减/畸形
- 盒底与男手掌接触合理
- 盒下间隙与桌面影子显示真实承托
Q6 平面与三维的同一空间 — 完整提示、四组目标与16项检查
原始提示 / Exact prompt
A single square architectural presentation divided into exactly two equal panels, one above the other, on a clean white page. The upper panel is titled 'PLAN' and is a true orthographic top-down floor plan. The lower panel is titled 'VIEW' and is an open-roof isometric rendering of the exact same small rectangular room. Both panels use the same shared object inventory and coordinate system. North is at the top of the plan and at the back of the isometric view; west is left and east is right. A small arrow labeled 'N' appears only in the plan.
The room contains exactly one round wooden table at its center, exactly three identical blue chairs, and exactly one potted green plant. Relative to the table, one chair is west, one is east, and one is south; there is no chair on the north side. The single plant is in the northeast corner of the room. There is one rectangular window centered in the north wall. There is one doorway in the west wall near the southwest corner; the door opening is not moved to the south wall. No other furniture, opening, rug or decorative plant is present.
The plan uses clear simple symbols and includes all three chair footprints, the circular tabletop, the northeast plant, the centered north window and the southwest portion of the west-wall doorway. The lower view uses believable wood, blue upholstery and neutral plaster. Choose an elevated viewing angle that makes the three chairs and both wall openings inspectable, with no roof hiding them. Maintain the same window, door, plant and chair positions across the two views. The only text is 'PLAN', 'VIEW' and the north arrow's 'N'.
考察范围: 两视图对象数量与方位、门窗位置、真实投影类型、北向对应和文字排除。
达标目标: 两图为同一空间,北侧无椅;家具及开口位置一致;下图应为轴测/等角图;只出现PLAN、VIEW、N。
双图结构(20分)
- 恰好上下2幅等大面板
- PLAN为正投影俯视
- VIEW为开顶轴测/等角
- 北在上图上方且在下图后方
共享数量(20分)
- 两图各1张圆桌
- 两图各恰好3椅且同为蓝色
- 两图各恰好1盆绿植
- 没有额外家具/开口/植物
椅与植物位置(20分)
- 两图椅子位于桌西/东/南
- 两图桌北无椅
- 两图植物在东北角
- 桌子均居中且大小关系可对应
门窗与文字(20分)
- 北墙居中窗两图一致
- 门在西墙近西南角两图一致
- 门未移到南墙且开口可检查
- 只出现PLAN/VIEW/N且标签对应
Q7 双参考局部修改与非目标保留 — 完整提示、四组目标与16项检查
原始提示 / Exact prompt
The red cube in the foreground of <Picture 1> is <Subject 1>. The uniform body color filling <Picture 2> is <Subject 2>. The three researchers, their faces, hair, clothing, poses and hands, together with their key and the water glass in <Picture 1>, are <Subject 3>. The green sphere, blue cylinder, yellow cone, magenta torus, presentation mat and glass table in <Picture 1> are <Subject 4>. The blackboard and all its existing lettering and chart, the rear mirror and its existing reflected content, walls, window, lighting and camera composition in <Picture 1> are <Subject 5>. The target image is the same laboratory photograph with only the red cube's body color changed to the exact donor color. <Subject 1> is partially preserved: its position, scale, orientation, cubic geometry, material and contact shadow remain, while its body color changes. <Subject 2> supplies the exact body color transferred to <Subject 1>. <Subject 3> is fully preserved. <Subject 4> is fully preserved. <Subject 5> is fully preserved. The transferred color follows the cube's existing highlights and shading, with physically necessary color changes to that cube's own reflection, without importing the flat donor canvas or changing any other object. The original square canvas, viewpoint and framing are retained. No person, object, label, pattern or panel is added or removed.
考察范围: 参考角色区分、供体色转移、局部性、人物身份与互动保留、非目标几何/文字/背景保留。
达标目标: 核心颜色编辑必须完成,其他区域应保留;不能把供体画布导入,或改脸/物/字。保留项高分不等于主编辑成功。
目标与供体(20分)
- 红方块改为参考2青绿色
- 改变的是方块而非球/柱/人物
- 保留原立方体位置尺度朝向
- 不把供体平面导入画面
人物与互动保留(20分)
- 仍为同三人脸型发型
- 服装及左右位置保留
- 手姿与钥匙交接保留
- 杯和吸管位置动作保留
非目标几何保留(20分)
- 绿球颜色形状位置保留
- 蓝柱/黄锥遮挡关系保留
- 洋红环形状颜色位置保留
- 垫子与桌子几何及视角保留
背景与限制(20分)
- 黑板原有字/数据/图保留
- 镜中原有内容保留不擅自修复
- 整体光照裁切构图保留
- 无额外人/物/字或删除
Q8 叙事、凝视与克制的真实感 — 完整提示、四组目标与16项检查
原始提示 / Exact prompt
A quiet, emotionally specific documentary photograph inside a tiny watch-repair shop on a rainy evening, square frame. Exactly two people occupy one coherent scene. On the left, an elderly woman watchmaker in a muted teal knitted cardigan sits at her worn wooden workbench. On the right, a young adult man in a damp brown wool coat leans toward her. One repaired silver pocket watch lies open in the woman's palm above the bench. It is the only watch or clock visible anywhere in the image.
The woman has paused before returning the watch. She looks at the young man's face, with the smallest restrained smile and a sense of relief. He looks down at the watch in her palm, not at the camera, with a subtle expression of recognition and gratitude. Neither person is broadly grinning or posing. Her free hand gently touches the edge of the watch without covering it. His hands remain together at the front edge of the bench; he has not taken the watch yet. The scene communicates a moment of shared history through these distinct gazes and the delayed handover, without visible speech or written explanation.
Only three small supporting tools occupy a narrow strip of the bench: a screwdriver, tweezers and a folded beige polishing cloth. Keep the rest of the work surface calm. A warm shaded task lamp at the left illuminates the hands, open metal watch and fine natural wrinkles. Cool diffuse rainy-window light from the rear right gives a soft rim to the damp coat. Rain beads, a few worn wood scratches, individual wool fibers and slight irregularities in skin add credible detail without filling the background with clutter. The background falls gently out of focus while both faces and the watch remain readable. Natural tonal range, controlled mixed lighting, tactile materials and honest human proportions; no glamour retouching, fantasy glow, extra person, extra watch, signage, caption or logo.
考察范围: 叙事时刻、不同凝视、克制情绪、手与怀表关系、三个工具计数、混合照明与材质真实感。
达标目标: 青年尚未接表;老妇看脸、青年看表;怀表及螺丝刀/镊子/米色布数量正确;左暖右冷光和材质服务画面焦点。
主体与排除(20分)
- 恰好两人及左老女/右青年男
- 青绿针织与棕色湿外套绑定
- 画面仅1只打开银怀表且无背景钟表
- 无额外人物/文字/标志
凝视与情绪(20分)
- 老妇看青年脸
- 青年看老妇掌中怀表
- 微笑克制非摆拍大笑
- 停顿与尚未交付的关系可读
手与道具(20分)
- 怀表在老妇掌中可见
- 老妇另一手轻触表边不遮表
- 青年双手在台前且尚未接表
- 仅螺丝刀/镊子/米色布三件支持道具
真实感与照明(20分)
- 左暖台灯与右后冷窗光有因果一致性
- 皮肤年龄/毛料/木料/金属均可信
- 雨珠湿衣与室内环境一致
- 细节服务焦点,背景克制不繁杂
所有题目及运行顺序的机器版本见exam.json,预提交哈希见exam-lock.json。Q3/Q4各两种子,其余各一种子;每路线10张,不能由这个样本量估计通用成功率。
本卷结果 / Results on these tasks only
| Test / 考核 | Six-step | pDMD two-step |
|---|---|---|
| Q1 Counting & attribute binding / 数量、属性绑定与空间遮挡 | 97.5 | 95 |
| Q2 Exact text & typography / 精确中文、拉丁字符与排版 | 100 | 97.5 |
| Q3 Quantitative charts / 图表的数据真实性 | 92.5 | 73.75 |
| Q4 Optics & reflection / 光学、镜像与接触物理 | 83.75 | 87.5 |
| Q5 Four-hand collaboration / 四只手的协作与归属 | 77.5 | 82.5 |
| Q6 Plan / 3D consistency / 平面与三维的同一空间 | 95 | 85 |
| Q7 Two-reference local editing / 双参考局部修改与非目标保留 | 87.5 | 80 |
| Q8 Narrative & gaze / 叙事、凝视与克制的真实感 | 97.5 | 100 |
| Macro average | 91.40625 | 87.65625 |
Q3 and Q4 use two preselected seeds; other tasks use one. Average seeds within each test first, then give each of the eight tests equal weight. Each recipe has ten images; the extra chart/optics seeds do not double those tests' weight.
Each image has 16 hard checks, five points each (0 / 2.5 / 5), plus coherence and aesthetics worth ten points each. Strict pass requires all 80 hard points. Both recipes strictly pass 3/10 images and 3/8 tests: counting, exact text and narrative. Overall hard scores are 75/80 (six-step) and 73.125/80 (two-step).
- Strong: exact Chinese and confusable Latin characters, count/attribute binding, photographic storytelling, differing gazes and restrained emotion.
- Brittle: chart geometry and table semantics, cropped mirror fields and refraction, exact hand-action stages, true axonometric projections.
- Editing trap: the six-step image scores 87.5 while leaving the red cube red. The two-step image changes its color, but misses the donor hue and repaints other content. Preservation points inflate the score; neither completes the intended local edit. The frozen v1 rubric is retained rather than fixed after seeing results.
What is included
data/: 20 untouched 4096×4096 RGB PNG outputs and 20 metadata rows.references/: the two exact Q7 inputs. The first is a synthetic laboratory image from the earlier six-step run; its existing mistakes are intentional preservation targets. The second is a uniform teal swatch.prompts/,exam.json,METHOD.md: full English prompts, exact visible copy, seeds and all Chinese atomic checks.workflows/: 20 actual submitted API graphs, including complete sampler/conditioning parameters.grades/,grade-lock.json,anonymous-assignment.json: original anonymous judgments, their frozen hashes and the later route mapping.exam-lock.json: pre-submission question/prompt/reference hashes.scores/,provenance/,diagnostics/: results, execution acceptance and supporting measurements. A corrected chart segmentation is explicitly retained beside the original diagnostic.gallery/: local HTML gallery, twenty previews and a contact sheet. A Dataset repository does not itself host this HTML as a Space; download it for offline viewing.SHA256SUMS: integrity manifest for every other file in this bundle;python verify.pychecks it using the standard library. The verifier excludes server-managed.gitattributesand local Git/HF client metadata when used in a repository checkout.
The original PNG bytes and embedded API prompt metadata are preserved. Their metadata was inspected for machine paths, identity and credential patterns before packaging. No model weights, private runtime logs, process snapshots, host paths, Git objects, global configuration or credentials are included. The official prompt-writing skill is credited and linked, not redistributed.
Evaluation procedure and limitations
Questions, scoring and seeds were frozen before generation. Every job uses its first successful output; there are no aesthetic retries, prompt changes, upscaling, refinement or generative touch-ups. Twenty anonymous visual grades were written and locked before route labels were revealed.
The grader and executor are the same AI agent. This is not independent review or a double-blind study. A score of 100 means the selected checks passed, not that the image is universally perfect. Only two seeds cover charts/optics; zero observed score range is not evidence of stability. These twenty synthetic examples do not establish production reliability or statistical significance.
Both recipes use the same INT8 ConvRot pruned base, NVFP4 AWQ text encoder and image VAE, on macOS arm64/MPS. The VAE has FP16 stored weights and FP32 runtime computation. This is not a full-BF16 ceiling test or a steps-only ablation. Exact graphs and component hashes are included; upstream model licenses remain with their respective publishers.
Recorded backend execution totals are 2028.430 seconds for six-step and 1020.559 seconds for two-step. Loading, task type, order and OS cache affect these measurements; they are not a controlled speed benchmark or evidence of lossless acceleration.
Dataset format and use
The card selects only data/** for the test split, so reference pictures and gallery previews are auxiliary assets rather than extra evaluation rows. data/metadata.jsonl follows the Hugging Face image dataset convention. The card format follows Dataset Cards. Local schema, file integrity and gallery links are validated; hosted Dataset Viewer conversion must be checked after an approved upload.
The metadata's prompt is the generation input; hard_checks holds the 16 grades in exam.json order. strict_pass is separate from total_score. reference_files holds relative paths for Q7 and an empty list for T2I cases. notes_zh contains the primary grader's observations. This is evaluation evidence rather than a curated training corpus.
引用本测试 / Citation
请连同任务范围、量化/运行条件和小样本限制引用,不把本卷成绩表述为LoRA总体能力排名。来源数据、题目、评分和工作流均可在本仓库检查。
@misc{h3_image_exam_v1_2026,
author = {edenessadzz},
title = {H3-Image Exam: Eight Tests, Twenty First-Try Images},
year = {2026},
url = {https://huggingface.co/datasets/edenessadzz/h3-image-exam-v1},
note = {Task-specific evaluation; quantized ComfyUI recipes on MPS; not an overall LoRA ranking}
}
Credits and license
MiniMax-H3-Image, the official prompt-writing skill, ComfyUI and the pDMD adapter publisher are upstream credits. No official endorsement or affiliation is claimed. Original prompts, rubric, annotations and generated examples in this release are offered under CC BY 4.0, to the extent applicable rights are held; third-party software, models and documentation retain their own licenses. Attribute this dataset's title and its eventual public repository URL.
中文速读
八科二十图,六步综合91.41、两步87.66,但两边都只有3/10图通过全部硬检查。中文、计数和叙事较强;图表、光学、精确动作和局部编辑会被漂亮画面掩盖。最关键的反例是六步编辑没改红块却拿87.5分。请先判断主目标是否完成,再看总分。完整逐项扣分、复测范围和原图见中文报告。本包是主agent自拟小样本评价,不是官方或独立排行榜。
- Downloads last month
- 444
