ACROSS-Lab/PromptTo3D_sd_dataset
收藏Hugging Face2024-07-08 更新2025-04-26 收录
下载链接:
https://hf-mirror.com/datasets/ACROSS-Lab/PromptTo3D_sd_dataset
下载链接
链接失效反馈官方服务:
资源简介:
---
dataset_info:
features:
- name: image
dtype: image
- name: label
dtype:
class_label:
names:
'0': train
- name: text
dtype: string
splits:
- name: train
num_bytes: 796865621.92
num_examples: 1030
download_size: 792357206
dataset_size: 796865621.92
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
---
# Stable Diffusion Dataset For 3D images generation
This is a set of 1.030 pairs prompt-image filtered and extracted from the 3D dataset [Objaverse](https://objaverse.allenai.org/)
This Dataset was used to finetune [Stable Diffusion 2](https://huggingface.co/stabilityai/stable-diffusion-2) in order to generate good (isolated, full object ...) images to feed an image to 3D model after that (like [Triposr](https://github.com/VAST-AI-Research/TripoSR) or [CRM](https://huggingface.co/Zhengyi/CRM)).
The issue we faced here was to chose which image to keep as most Objaverse objects had no title, no description or an inadequate one (for instance : School Project n°45). Thus the images have been sorted manually and by keeping around 20% of them we managed to build a 1000 big image dataset.
数据集信息:
特征:
- 名称:图像(image),数据类型:图像
- 名称:标签(label),数据类型:
分类标签(class_label):
类别名称:
'0': 训练集(train)
- 名称:文本(text),数据类型:字符串
数据划分:
- 名称:训练集(train),字节占用数:796865621.92,样本数量:1030
下载大小:792357206,数据集总大小:796865621.92
配置项:
- 配置名称:默认(default),数据文件:
- 划分集:训练集(train),路径:data/train-*
---
# 面向3D图像生成的Stable Diffusion数据集
本数据集包含1030条提示词-图像配对样本,均从3D数据集[Objaverse](https://objaverse.allenai.org/)中筛选并提取而来。
本数据集曾用于微调[Stable Diffusion 2](https://huggingface.co/stabilityai/stable-diffusion-2),以生成高质量(孤立背景、完整物体等)图像,后续可将这些图像输入至3D模型(如[Triposr](https://github.com/VAST-AI-Research/TripoSR)或[CRM](https://huggingface.co/Zhengyi/CRM))中。
本次研究面临的核心问题是如何筛选有效图像,因为Objaverse中的绝大多数物体既无标题也无描述,或其描述并不适配生成需求(例如“学校项目第45号”)。因此,研究团队通过人工筛选的方式对图像进行排序,最终保留了约20%的样本,成功构建出包含1000条样本的图像数据集。
提供机构:
ACROSS-Lab



