@YeheLiu@itsmetamike Looks great! Are you using our InfiniSplat v3 (https://t.co/MRw4robaaW)? If you’re using a different model, would you mind sharing which one?
The InfiniSplat team have released v3 as SplatAnything. Just 22 days between the v2 and v3 papers 🤯
V2 could already turn overlapping photos into a Gaussian splat without training each scene from scratch. In v3, the team uses the spacing between splats to help decide how big they should be. Too small and you see gaps, while too much overlap softens the detail.
V3 is designed to get that balance right whether you’re scanning an apple on a desk or a whole city block.
A tabletop. A room. A city.
SplatAnything reconstructs them with the same feed-forward 3D model. The key is normalizing Gaussian spacing, rather than scene size alone.
Here's the idea in 3 minutes.
#3DGS#ComputerVision
Thrilled to release InfiniSplat v3 (SplatAnything)!
We teamed up with @TencentHunyuan to scale up InfiniSplat v2’s training data & introduced Gaussian Spacing Normalization for high-quality rendering across all scales, from object to aerial.
Code is now open-source👇
@YeheLiu Thank you for your recognition and support of our work! We will continue to actively develop and maintain the InfiniSplat series, striving to advance the field of feed-forward 3D Gaussian reconstruction.
Excited to share SplatAnything!
From everyday objects and human subjects to expansive aerial scenes, SplatAnything transforms simple captures into 3D Gaussian scenes in few seconds. Give it a Try!
🌐 Project Page: https://t.co/1A0EmjsFQ8
🤗 Huggingface: https://t.co/csbGtLtLNU
@YeheLiu Sorry for the confusion. There was an error in our previous license, but we have now updated it with a corrected version. Please feel free to check out the new license!
Excited to share SplatAnything!
From everyday objects and human subjects to expansive aerial scenes, SplatAnything transforms simple captures into 3D Gaussian scenes in few seconds. Give it a Try!
🌐 Project Page: https://t.co/1A0EmjsFQ8
🤗 Huggingface: https://t.co/csbGtLtLNU
Fable 5.2 shows substantially improved skills in controlling physical robots, even on tasks it has never explicitly been trained to do.
On RoboDojo, a benchmark with a simulated two-arm robot, it outperformed other models and most AI systems built to control robots.
@zoom_will Thanks! Actually, SplatAnything supports input images at various resolutions, including 1K and 2K. Feel free to try it with higher-resolution images!
SplatAnything
Gaussian Spacing Normalization for feed-forward 3DGS across arbitrary scenes.
From objects and people to aerial views, this method reconstructs 3D Gaussian scenes from simple captures with state-of-the-art quality.