PixGR: pixel-space generative decoding for fast feed-forward 3D Gaussian reconstruction
zhenjun_zhao · x · 2026-10-27
New arXiv paper PixGR (2610.31908; Huang, Bian, Luo, Mao, Mikolajczyk) integrates lightweight pixel-space generative modeling into feed-forward 3D Gaussian reconstruction to model scene uncertainty — tl;dr: VGGT + DiT decoder.
Method: context views are encoded once into a scene latent; a lightweight DiT decoder (global self-attention + Gaussian head) generatively decodes noisy target views into a shared set of pixel-aligned 3D Gaussians. Formulated with pixel-space flow matching and optimized end-to-end via differentiable rendering, avoiding VAE latent compression and error propagation of two-stage pipelines. Scene latents are cached, keeping inference sub-second.
Results:
- View interpolation: PSNR 28.54 / SSIM 0.905 on RE10K, 25.27 on DL3DV — beating regression and generative baselines like DepthSplat and ReconSplat;
- View extrapolation: clear wins, sharper and more geometrically consistent renders without the artifacts/misalignment of generative baselines;
- Biggest gains in challenging sparse-view settings with widely spaced observations, while staying lighter and faster than heavyweight pretrained generative models.
More from Research
- UK AISI: under $80 of fine-tuning fully restores filtered bio capabilities in open-weight models — connoraxiotes · 2026-10-27
- RFD4-Proteina runs on an RTX 3060: protein design from $0.001 per design — AllThingsApx · 2026-10-27
- University of Freiburg's Generative Intelligence Lab recruiting funded PhDs via ELLIS, apply by Oct 31 — yuyinzhou_cs · 2026-10-27
- Nature Physics: reprogrammable refractive index enables arbitrary multimode wave control for ML — MarioKrenn6240 · 2026-10-27
- Wakaru open-sources ontologize: labeling datasets from model activations — amplifiedamp · 2026-10-27
- EMNLP Oral paper traces multilingual representations in LLMs with cross-layer transcoders — ZhijingJin · 2026-10-27