Copy README.md from OmniWeave/Consis3D-Bench (verbatim)

README.md

@@ -0,0 +1,74 @@
+---
+license: apache-2.0
+language:
+- en
+task_categories:
+- image-to-3d
+- text-to-3d
+- other
+tags:
+- 3d
+- 3d-editing
+- mesh
+- benchmark
+size_categories:
+- n<1K
+pretty_name: Consis3D-Bench
+---
+
+<h1 align="center" style="font-size: 3em; margin-bottom: 0.2em;">Consis3D-Bench</h1>
+
+The evaluation benchmark of the paper
+**[Omni123: Exploring 3D Native Foundation Models with Limited 3D Data by Unifying Text to 2D and 3D Generation](https://arxiv.org/abs/2604.02289)**.
+Consis3D-Bench is the **first 3D editing benchmark with meshes sculpted by professional artists**.
+
+## Source of the data
+
+The pipeline is the same as [3DEdit-1M](https://huggingface.co/datasets/OmniWeave/3DEdit-1M), except that the target 3D meshes are sculpted by professional artists:
+
+- **Edit instruction** synthesized by [Qwen3.5-35B-A3B](https://huggingface.co/Qwen/Qwen3.5-35B-A3B).
+- **Target image** generated by [FLUX.2-klein-9b-kv](https://huggingface.co/black-forest-labs/FLUX.2-klein-9b-kv).
+- **Source 3D meshes** generated by [Hunyuan3D-2-mini](https://github.com/Tencent-Hunyuan/Hunyuan3D-2) from the source images.
+- **Target 3D meshes** hand-sculpted by **professional artists** from the source mesh, following the edit instruction and target image.
+
+## Dataset structure
+
+```
+benchmark/
+├── <uuid>/
+│   ├── source.glb
+│   ├── target.glb
+│   ├── consistent_mesh.glb
+│   ├── source.rgba.webp
+│   ├── target.rgba.webp
+│   ├── normal_front.png
+│   ├── normal_back.png
+│   ├── normal_left.png
+│   ├── normal_right.png
+│   ├── normal_top.png
+│   ├── normal_bottom.png
+│   └── instruction.txt
+└── ...
+```
+
+| File | Description |
+| --- | --- |
+| `source.glb` | Source mesh from Hunyuan3D-2-mini, normalized so the longest bbox extent fits in `[-1, 1]³` |
+| `target.glb` | Artist-edited target mesh (same frame as source) |
+| `consistent_mesh.glb` | The unchanged region of `source.glb` (used to score consistency) |
+| `source.rgba.webp` | 4-view fused source RGBA render |
+| `target.rgba.webp` | 4-view fused target RGBA render |
+| `normal_{front,back,left,right,top,bottom}.png` | 512×512 view-space normal maps of the target mesh |
+| `instruction.txt` | Edit instruction synthesized by [Qwen3.5-35B-A3B](https://huggingface.co/Qwen/Qwen3.5-35B-A3B) |
+
+All meshes share the same coordinate frame: Y-up, centered at the origin, target's longest bounding-box extent normalized to 2.0 (mesh fits inside `[-1, 1]³`).
+
+## Evaluation Metrics
+
+1. **Consistent average Chamfer distance** — one-way CD from the unchanged region of source to the prediction.
+2. **ULIP-2 point feature similarity** — cosine similarity between gt and pred ULIP-2 point embeddings.
+3. **6-view normal image similarity** — DINO-I / CLIP-I / SSIM / LPIPS between `normal_*.png` and the corresponding renders of the prediction.
+
+## License
+
+Released under the **[Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0)**.