DNA-Rendering

该数据集专为用于定量比较在人类形象生成中的扩散模型而设计，包含了多视角的人类动作视频。该数据集被用于评估重建指标（如PSNR、SSIM和LPIPS），生成质量指标（如FID和Patch-FID），以及跨视角一致性（如PPLC）。其规模达到百万像素分辨率（1024x1024），任务定位于在百万像素分辨率下进行人类图像生成。

This dataset is specifically designed for the quantitative comparison of diffusion models used for human figure generation. It includes multi-view human action videos. This dataset is employed to evaluate reconstruction metrics including PSNR, SSIM and LPIPS, generation quality metrics including FID and Patch-FID, as well as cross-view consistency metrics including PPLC. It features a resolution of 1024×1024 pixels (roughly 1 megapixel), and the task is oriented towards human image generation at megapixel resolution.

搜集汇总

背景与挑战

背景概述

DNA-Rendering数据集专为人类形象生成中扩散模型的定量比较而设计，包含多视角人类动作视频，用于评估重建、生成质量和跨视角一致性指标。其规模达到百万像素分辨率（1024x1024），核心任务是在高分辨率下进行人类图像生成评估。