dylanebert/iso3d

---
license: mit
---

# iso3d - Isolated Synthetic Objects 3D

A dataset of isolated object images for evaluating [image-to-3D](https://huggingface.co/models?pipeline_tag=image-to-3d) models.

## Leaderboard

Vote and view results at [3d-arena](https://huggingface.co/spaces/dylanebert/3d-arena).

## Curation

Images are created using [dreamshaper-xl](https://huggingface.co/Lykon/dreamshaper-xl-v2-turbo) and [white background lora](https://civitai.com/models/119388/white-background) on [karlo-v1](https://huggingface.co/datasets/diffusers-parti-prompts/karlo-v1) prompts.

1. Each [karlo-v1](https://huggingface.co/datasets/diffusers-parti-prompts/karlo-v1) prompt is extended with `{prompt}, isolated object render, with a white background` and negative prompt `text, watermark, shadow, background`.
2. Images are generated using [ComfyUI](https://github.com/comfyanonymous/ComfyUI) with [dreamshaper-xl](https://huggingface.co/Lykon/dreamshaper-xl-v2-turbo) and [white background lora](https://civitai.com/models/119388/white-background).
3. Backgrounds are removed using [rembg](https://github.com/danielgatis/rembg).
4. 100 images are hand-selected from the 1.63k generated images.

## Contributing

The leaderboard is automatically populated by the [3d-arena dataset](https://huggingface.co/datasets/dylanebert/3d-arena).

To submit your model, [open a PR](https://huggingface.co/docs/hub/en/repositories-pull-requests-discussions) on the dataset.

## Citation

```
@misc{3d-arena,
author = {Dylan Ebert}
title = {3D Arena}
year = {2024}
publisher = {Hugging Face}
howpublished = \url{https://huggingface.co/spaces/dylanebert/3d-arena}
}
```

原始信息汇总

iso3d - Isolated Synthetic Objects 3D

数据集描述

iso3d数据集包含用于评估image-to-3D模型的孤立物体图像。

数据集制作

1. 使用karlo-v1提示词，扩展为{prompt}, isolated object render, with a white background，并添加负提示词text, watermark, shadow, background。
2. 通过ComfyUI生成图像，使用dreamshaper-xl和white background lora。
3. 使用rembg去除背景。
4. 从生成的1.63k图像中手动挑选100张图像。

引用

@misc{3d-arena, author = {Dylan Ebert}, title = {3D Arena}, year = {2024}, publisher = {Hugging Face}, howpublished = url{https://huggingface.co/spaces/dylanebert/3d-arena} }