Datasets:
|
Download README.md from yomi-chen/EditableDesignBench_result: direct link, hf CLI and curl.
- Browser
- Download file 6.7 kB
-
https://huggingface.co/datasets/yomi-chen/EditableDesignBench_result/resolve/main/README.md
- Command line
-
hf download hf://datasets/yomi-chen/EditableDesignBench_result/README.md
-
curl -L -o README.md https://huggingface.co/datasets/yomi-chen/EditableDesignBench_result/resolve/main/README.md
6.7 kB
| pretty_name: EditableDesignBench Run3 Results | |
| language: | |
| - en | |
| - zh | |
| license: cc-by-nc-4.0 | |
| task_categories: | |
| - image-to-image | |
| tags: | |
| - image-editing | |
| - benchmark | |
| - generated-results | |
| - EditableDesignBench | |
| size_categories: | |
| - 1K<n<10K | |
| # 🎨 EditableDesignBench_result | |
| EditableDesignBench **v4.0.0 / run3** 的六个模型结果:每模型 **文字 100 张 + 图片资产 100 张**,共 **1,200 张**。 | |
| 全部图片与 MSE、PSNR、SSIM、LPIPS 已完成。**尚未进行六个模型的全量 MLLM 评分。** | |
| [数据集](https://huggingface.co/datasets/yomi-chen/EditableDesignBench) · [评测代码](https://github.com/Ch3nSir/EditableDesignBench) | |
| ## 📥 下载并解压 | |
| ```bash | |
| pip install -U huggingface_hub | |
| hf download yomi-chen/EditableDesignBench_result --repo-type dataset --local-dir EditableDesignBench_result | |
| python EditableDesignBench_result/extract.py | |
| cd EditableDesignBench_result | |
| ``` | |
| 公开下载,无需登录。也可在 Files 页面只下载需要的模型 ZIP;同时下载 `EditableDesignBench_v4_data.zip`,即可取得对应原图、指令和评测标注。 单独下载时,运行评测还需下载 `evaluation_code.zip`。 | |
| | 模型 | 文字 | 资产 | ZIP | | |
| |---|---:|---:|---| | |
| | Fibo-2.0-edit-turbo | 100 | 100 | `archives/fibo2-edit-turbo_run3.zip` | | |
| | JoyAI-Image-Edit-Plus | 100 | 100 | `archives/joyai-image-edit-plus_run3.zip` | | |
| | Qwen-Image-2.1 | 100 | 100 | `archives/qwen-image21_run3.zip` | | |
| | Qwen-Image-2.1-Turbo | 100 | 100 | `archives/qwen-image21-turbo_run3.zip` | | |
| | FLUX 3 Image Klein | 100 | 100 | `archives/flux3-klein_run3.zip` | | |
| | SenseNova-U1.5-8B-MoT | 100 | 100 | `archives/sensenova-u1.5-8b-mot_run3.zip` | | |
| ## 📂 解压后的内容 | |
| ```text | |
| data/EditableDesignBench-v4/ # 本次使用的完整 200 条任务:原图、指令、mask、目标图 | |
| text/<model>/predictions/<id>.png # 文字结果的 RGB 评测图片 | |
| assets/<model>/predictions/<id>.png # 资产结果的 RGB 评测图片 | |
| text/<model>/raw_outputs/<id>.png # 模型原始返回文件,逐字节保留 | |
| assets/<model>/raw_outputs/<id>.png | |
| text/<model>/manifest.jsonl # 逐例指令、路径、尺寸及生成信息 | |
| assets/<model>/manifest.jsonl | |
| metrics/summary.csv # 两轨道的模型像素指标汇总 | |
| metrics/case_pixel_scores.csv # 1,200 条逐例像素指标 | |
| predictions_index.jsonl # 全部 1,200 条任务与图片的对应关系 | |
| release_info.json # 数据、代码版本和模型配置 | |
| ``` | |
| `predictions_index.jsonl` 中的路径都相对于本目录,每条记录包含 `source`、`prediction`、`raw_output`、`instruction`、`model`、`track`、`operation` 和尺寸信息。 | |
| `raw_outputs/` 保留全部原始输出,包括可能存在的 alpha 通道。其余五个模型的 `predictions/` 仅取原始 RGB 通道并去掉已记录的右侧/底部 padding,**不缩放、不合成背景、不重绘**。JoyAI 的 `predictions/` 保留官方分辨率的完整 RGB 输出。 | |
| 模型 ZIP 中的 `recorded_reports/` 为本次运行的历史生成记录及像素报告,原机器路径已替换成逻辑路径标签,仅供溯源。跨机器评测请使用上述 `predictions/` 与 `manifest.jsonl`,不要直接用历史报告恢复缓存。 | |
| ## 🧪 使用现有图片做 MLLM 评分 | |
| 结果包附带本次代码快照 `archives/evaluation_code.zip`,运行 `extract.py` 时会一起解压。无需下载模型权重或重新生成图片。 已在 Python 3.12 环境验证,纯 API 评分不需要显卡。 | |
| 下面以 Fibo 为例;六个模型都可使用同一命令,只需替换模型目录。MLLM 允许原图和结果图尺寸不同,两张完整图片都保持原样,不缩放或裁剪: | |
| ```bash | |
| pip install -r evaluation_code/requirements.txt | |
| export OPENAI_API_KEY="your_api_key" | |
| export OPENAI_API_URL="https://your-provider.example/v1" | |
| export OPENAI_MODEL="your-vision-model" | |
| # 文字 100 条 | |
| python evaluation_code/scripts/mllm_evaluate.py \ | |
| --dataset data/EditableDesignBench-v4 \ | |
| --predictions text/fibo2-edit-turbo/predictions \ | |
| --output judge/text/fibo2-edit-turbo.json \ | |
| --workers 10 --require-scoring | |
| # 资产 100 条 | |
| python evaluation_code/scripts/mllm_evaluate.py \ | |
| --dataset data/EditableDesignBench-v4/assets \ | |
| --predictions assets/fibo2-edit-turbo/predictions \ | |
| --output judge/assets/fibo2-edit-turbo.json \ | |
| --workers 10 --require-scoring | |
| ``` | |
| 输入为完整原图、完整结果图和自然语言指令,不提供 mask 或目标图给评委。每例一次请求返回四维评分;重复命令会复用已成功的评分,只补失败/缺失项。上述命令会实际调用使用者配置的评委 API;发布的模型图片不需要重新生成。 遇到接口限流或超时时,可重跑同一命令并将 `--workers 10` 改为 `--workers 2`,已成功的维度不会重复请求。 | |
| ### JoyAI:MLLM 与像素指标的尺寸规则不同 | |
| **MLLM 可直接接收不同尺寸的两张整图。** JoyAI 使用官方默认档位(1024-base bucket,30 steps,CFG 4);例如原图 1600 × 1200、输出 1152 × 832,评委接收两张各自完整的图与指令,不缩放、不裁剪、不补边、不进行 Alpha 合成: | |
| ```bash | |
| python evaluation_code/scripts/mllm_evaluate.py \ | |
| --dataset data/EditableDesignBench-v4 \ | |
| --predictions text/joyai-image-edit-plus/predictions \ | |
| --output judge/text/joyai-image-edit-plus.json \ | |
| --workers 10 --require-scoring | |
| ``` | |
| 资产轨同时将数据目录改为 `data/EditableDesignBench-v4/assets`、图片目录改为 `assets/joyai-image-edit-plus/predictions`。不同尺寸不会直接导致拒收;内容丢失或布局变化仍按原有 rubric 正常评分。 | |
| **像素指标需要对应的评测坐标。** 已发布的 JoyAI 像素分数将原图和 mask 按官方比例缩放、中心裁剪后,与未缩放的结果图比较。变换记录在 `official_geometry` 中;被裁掉的边缘不参与像素计算。因此它与其他模型的原尺寸像素分数应分开解读。 | |
| 已下载旧包的用户只需重新下载 `archives/evaluation_code.zip`,解压覆盖 `evaluation_code/`;图片与既有像素指标均未改变。 | |
| ## 📊 指标与版本 | |
| - 文字和资产分开统计;四项像素指标均覆盖各模型全部 200 条任务。 | |
| - 像素指标衡量非编辑区域的保持程度,不代表编辑任务已完成。 | |
| - 数据版本:EditableDesignBench `v4.0.0`;生成 seed:每例 `42`。 | |
| - 本次代码版本见 `release_info.json`,评测 prompt 与代码快照一致。 | |
| - 数据集沿用 CC BY-NC 4.0;代码快照中的代码许可证与第三方声明见其 `LICENSE`、`NOTICE.md`。各模型及组件另有各自的上游条款。 | |