--- license: other pretty_name: SHARP:单张图的秒级 3D 高斯视图合成 tags: - 3D vision - gaussian splatting - view synthesis - Apple --- # SHARP:单张图的秒级 3D 高斯视图合成 SHARP 是 Apple 提出的单目视图合成方法:给定一张照片,它在单次前馈中回归出场景的 3D 高斯表示,标准 GPU 上耗时不到 1 秒。得到的 3DGS 可以被实时渲染,输出附近视角的高分辨率写实图像,而且表示是 metric 的,带绝对尺度,支持米制相机运动。 按 README 与论文摘要,它在多个数据集上把 LPIPS 降低 25–34%、DISTS 降低 21–43%,同时把合成耗时降低三个数量级。 对开发者来说,值得看的是它把「重建」压缩成了一次前馈调用,并且给出了现成的 CLI 与权重。 This repository collects reading notes and links. **No model weights are redistributed here** — the official checkpoint and code stay with their authors; follow the links below. ## Key facts - **合成耗时**: 标准 GPU 上不到 1 秒,单次前馈 - **LPIPS**: 较此前最佳模型降低 25–34% - **DISTS**: 较此前最佳模型降低 21–43% - **加速幅度**: 合成耗时降低三个数量级 - **输出表示**: metric 3D 高斯(3DGS),输出 .ply 文件 - **泛化**: 论文称具备鲁棒的零样本泛化 ## Sources - SHARP 通过单次前馈在标准 GPU 上以不到 1 秒的时间回归出场景的 3D 高斯表示参数。 — [raw.githubusercontent.com](https://raw.githubusercontent.com/apple/ml-sharp/main/README.md) - SHARP 产出的 3D 高斯表示可以被实时渲染,生成附近视角的高分辨率写实图像。 — [raw.githubusercontent.com](https://raw.githubusercontent.com/apple/ml-sharp/main/README.md) - 该表示是 metric 的,带绝对尺度,支持米制的相机运动。 — [raw.githubusercontent.com](https://raw.githubusercontent.com/apple/ml-sharp/main/README.md) - 实验结果展示了 SHARP 跨数据集的鲁棒零样本泛化能力。 — [raw.githubusercontent.com](https://raw.githubusercontent.com/apple/ml-sharp/main/README.md) - 在多个数据集上,SHARP 相对此前最佳模型把 LPIPS 降低 25–34%、DISTS 降低 21–43%,同时把合成耗时降低三个数量级。 — [raw.githubusercontent.com](https://raw.githubusercontent.com/apple/ml-sharp/main/README.md) - 预测结果是以 3D 高斯泼溅(3DGS)形式存放在输出目录中的可渲染结果。 — [raw.githubusercontent.com](https://raw.githubusercontent.com/apple/ml-sharp/main/README.md) - 官方 CLI 的预测入口是 sharp predict,用 -i 指定输入图像目录、-o 指定输出高斯目录。 — [raw.githubusercontent.com](https://raw.githubusercontent.com/apple/ml-sharp/main/README.md) - 模型权重在第一次运行时自动下载,并缓存到 ~/.cache/torch/hub/checkpoints/。 — [raw.githubusercontent.com](https://raw.githubusercontent.com/apple/ml-sharp/main/README.md) - 高斯预测在 CPU、CUDA 和 MPS 上都能工作,但通过 --render 渲染相机轨迹视频目前需要 CUDA GPU。 — [raw.githubusercontent.com](https://raw.githubusercontent.com/apple/ml-sharp/main/README.md) - 输出的 3DGS 遵循 OpenCV 坐标约定(x 右、y 下、z 前),场景中心大致在 (0, 0, +z),接第三方渲染器时需要自行缩放与旋转来重新居中。 — [raw.githubusercontent.com](https://raw.githubusercontent.com/apple/ml-sharp/main/README.md) ## Links - [arXiv 2512.10685](https://arxiv.org/abs/2512.10685) - [项目主页](https://apple.github.io/ml-sharp/) - [官方仓库](https://github.com/apple/ml-sharp) - [模型权重](https://ml-site.cdn-apple.com/models/sharp/sharp_2572gikvuh.pt) ## More SHARP 展示了一件值得记住的事:一张普通照片里,本来就藏着足够多的信息,只要有一个足够鲁棒的表示方式把它读出来。 在你自己的照片上,最实用的第一步往往不是重建 3D,而是先把图像本身的质量拉起来。 DLSS 5 是一个在线 AI 图像编辑器,把手里已有的图画上传上去,就能试试我们的 AI 超分与画质增强。 [上传图片,试试 AI 增强](https://www.dlss5nvidia.com) --- Notes compiled from `topics/ml-sharp/` with content-engine.