--- title: GaussianRenderFormer emoji: ✨ colorFrom: pink colorTo: blue sdk: gradio sdk_version: 6.31.0 app_file: app.py python_version: "3.12" short_description: Transformer-based rendering of 3D Gaussian splats startup_duration_timeout: 30m --- # GaussianRenderFormer **Transformer-based rendering of 3D Gaussian splats** — pick a Gaussian Splatting scene, point an orbit camera at it, and a transformer renders the view. No rasterizer, no per-scene optimization: each Gaussian's 14 parameters become one scene token, a view decoder cross-attends over 4-px image patches, and a DPT head produces the image. - Paper: [arXiv:2611.05593](https://arxiv.org/abs/2611.05593) — Valencia Levy & Benaim, 2026 - Code: [github.com/SVLwoof/gaussianformer](https://github.com/SVLwoof/gaussianformer) (MIT) - Checkpoint: [shahafvl/gaussianformer](https://huggingface.co/shahafvl/gaussianformer) (MIT) ## What you can do here - **Example scenes** — three synthetic splats authored for this demo (a faceted gem, a glazed vase, a ringed planet, each ≤ 20k Gaussians) rendered through the exact PLY → `read_ply` → `normalize` path the model uses in production. - **Your own splat** — upload a 3DGS `.ply` (raw optimizer fields: log scales, wxyz quaternions, SH colors, opacity logits). Tell the app the file's up axis and it centers, re-orients and scales the object into the model's frame. Splats larger than 20k Gaussians are pruned by an importance score; the model was trained on ~20k-Gaussian objects and denser inputs lose quality. The camera orbits the object's center: azimuth, elevation and distance (the model was trained at distances 1.15–2.45, 45° FOV, 512 px). ## Deviations from the reference pipeline The reference `load_ply` prunes with gsplat-projected radii and then fine-tunes the kept Gaussians against the full splat for ~1500 iterations. A CUDA rasterizer with a JIT build does not fit this demo's runtime, so pruning here uses an analytic approximation of the same importance score (opacity × projected area estimated from pinhole geometry, × scale^0.1) without the fine-tune. Per the model card, large pruned splats can "lose coverage and turn dark" without the fine-tune — the bundled examples are ≤ 20k Gaussians and so are never pruned. ## Example scenes & licensing The three example scenes are original synthetic splats generated procedurally for this demo (no captured data), so they carry no third-party license restrictions. The model was trained on [Objaverse_Splats](https://huggingface.co/datasets/ShapeSplats/Objaverse_Splats), whose objects each keep their creator's license — that is why the demo does not bundle any of the paper's evaluation scenes. ## Limitations (from the model card) Trained on isolated single objects centered, Y-up and scaled into [-0.45, 0.45]³, at 512 px and 45° FOV; denser inputs lose quality; colors are view-independent (only each Gaussian's DC color is used); fine detail renders softer than a rasterizer.