Spaces:
Running on Zero
Running on Zero
File size: 3,010 Bytes
655c00b 4e72893 655c00b 4e72893 655c00b 4e72893 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 | ---
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.
|