license: mit
tags:
- 3d-gaussian-splatting
- novel-view-synthesis
- pose-free
- feed-forward
datasets:
- re10k
pipeline_tag: image-to-3d
EvenSplat: Compact Pose-Free Feed-Forward 3D Gaussian Splatting
Kyoosuk Hwang · Junyoung Kwon · Seoul National University of Science and Technology · Capstone Design 2026
EvenSplat renders novel views from two unposed images using 25 to 75% of the per-pixel Gaussians that NoPoSplat predicts. It keeps the same fraction of Gaussians in every 2 × 2 pixel cell, drops low-reliability Gaussians first, and is fine-tuned with that pruning in the loop.
Checkpoints
| File | Gaussians kept | #Gaussians | PSNR ↑ | SSIM ↑ | LPIPS ↓ |
|---|---|---|---|---|---|
retention_r25_re10k.ckpt |
25% | 33K | 24.62 | 0.817 | 0.184 |
retention_r50_re10k.ckpt |
50% | 66K | 25.01 | 0.829 | 0.169 |
retention_r75_re10k.ckpt |
75% | 98K | 25.26 | 0.837 | 0.161 |
| NoPoSplat (reported) | 100% | 131K | 25.03 | 0.838 | 0.160 |
RealEstate10K full test set (5,590 evaluable scenes), 256 × 256, two context views, pose alignment on. At 50%, peak render memory drops from 167 MB to 87 MB.
Each file holds model weights only (no optimizer state), about 2.45 GB. They are for evaluation; to fine-tune
again, start from the base checkpoint
kyoosuk/noposplat-olp-rel-re10k
(epoch_1-step_75000.ckpt). Every model was fine-tuned from it for 10K steps at its own ratio: 2 × 2 cells,
r threshold 0.35, batch 1 × 4 accumulation, lr 2e-5.
Usage
Set up the environment and data as in the repository README, then:
hf download kyoosuk/EvenSplat retention_r50_re10k.ckpt --local-dir checkpoints
python scripts/run_retention.py --config my_settings.yaml eval-grid --split full \
--checkpoint checkpoints/retention_r50_re10k.ckpt \
--baselines none --scores r_filter --ratios 0.5 --run-id ours_r50 --execute
Use the --ratios value that matches the file (0.25, 0.5 or 0.75).
Acknowledgements
Built on NoPoSplat, and in turn on pixelSplat, DUSt3R and CroCo. Released under the MIT License.