Skip to main content
Search

# Gaussian Splatting

13 related articles

Subscribe to updates
  1. arXiv

    CrossPressor sidesteps the fusion bottleneck in long-sequence 3D reconstruction: training-free routing cuts 256-view cross-view fusion from 125 ms to 4.8 ms

    As feed-forward 3D Gaussian splatting scales to more views, anchor-based compression fixes the decoder-facing representation but lets support-to-anchor fusion grow with view count. CrossPressor leaves released checkpoints untouched and routes support features before K/V projection with coverage-cycled offsets, spreading retained spatial evidence across support views and fusion layers. At 256 views it keeps 1.85% of dense QK interactions, cutting cross-view latency from 125.28 ms to 4.80 ms while rendering quality stays close to dense fusion.

  2. arXiv

    Gaussian splats skip the rasterizer: GaussianRenderFormer renders directly, reaching 38.08 dB on unseen objects

    GaussianRenderFormer replaces the hand-written rasterizer in 3D Gaussian splatting with a transformer: one token per Gaussian, and a view-dependent decoder that turns small patches of camera rays into pixels. Trained on 26,820 Objaverse objects, it reaches 38.08 dB PSNR on 300 unseen objects against renders of the full splat, while rasterizing the same pruned input reaches 44.02 dB; diagnostics point to a view-decoder bottleneck, though under-training is not ruled out.

  3. arXiv

    One photo is enough for a 3D plant: Plant Orbiter orbits a video diffusion model around the specimen and holds mean height error to 5.9%

    Digitizing a plant in 3D usually takes tens to hundreds of calibrated photographs, or costly laser scanning and robotic capture, while single-image generators struggle to align their output back to the input photo. Plant Orbiter uses a video diffusion model to synthesize an orbit video around the plant from one frontal image, then fits 3D Gaussian splats using the known camera path, skipping structure-from-motion camera recovery. Across PlantDreamer, Botany-Bot and PLANesT, its 3D F-score leads the compared methods on every tested dataset, its input-view IoU reaches 0.552 and 0.556 on PlantDreamer and Botany-Bot, and height measured on the GrowSplat time series gives a mean absolute percentage error of 5.9%.

  4. Trained Gaussian-splat scenes no longer need retraining to shrink: post-hoc entropy coding reaches 6.0x on 13 scenes at a 0.03 dB quality change

    Trained anchor-based Gaussian-splat scenes stay too large for phones and headsets, and existing post-hoc converters saturate early. This scheme treats a trained model as a read-only artifact, fits a per-scene hash-grid entropy context with per-channel Laplace distributions, and codes attributes into a Morton-ordered, progressively decodable bitstream. Across 13 scenes on three benchmarks, the post-hoc mode averages 6.0x compression at a -0.03 dB quality change and runs unchanged on third-party checkpoints.

  5. arXiv

    World models stop growing with video length: L3G-World keeps a latent 3D Gaussian memory and runs at 9.02 FPS on one H100

    Interactive world models have been limited by memory that grows with stream length: temporal context runs out of GPU memory, and explicit 3D memories slow down as redundant primitives accumulate frame by frame. L3G-World replaces that with a compact latent 3D Gaussian memory that spawns Gaussians only in newly observed regions, refines existing ones with residual updates, and stores a VAE latent feature in each Gaussian so the memory renders directly into the diffusion transformer's latent space without re-encoding. On 481-frame loop-closure trajectories from DL3DV and RealEstate10K, full L3G-World reaches PSNR of 16.00 and 18.36, and with few-step distillation it generates at 9.02 FPS on a single H100.

  6. arXiv

    3D Gaussian Splatting research without hand-written code: SplatLab runs from a one-line prompt to a manuscript, each beating its reproduced baseline

    SplatLab hands the 3D Gaussian Splatting research loop to 36 agents: it reproduces a published baseline, mutates a verified plugin, trains, ablates and writes a manuscript, binding every reported number to a logged GPU run. On Mip-NeRF 360 and LLFF its manuscripts beat their reproduced targets on PSNR; on astrophotographic nebula rendering, where no prior 3DGS method exists, Nebula3DGS-v2 raises PSNR from the 3DGS baseline's 32.58 dB to 36.36 dB.

  7. arXiv

    Driving planners get the missing future frames: rendering along candidate trajectories lifts NAVSIM v1 from 86.2 to 91.3

    Driving world action models explore candidate trajectories absent from logs during reinforcement-learning post-training, but logs store future video only for recorded trajectories, so those candidates lack matched visual supervision. SimVTA renders future video along candidate paths with a simulator that reconstructs real scenes via 3D Gaussian Splatting, giving prediction-independent targets, then selects high-reward, high-error futures by trajectory reward and video prediction error for flow-matching supervision while GRPO optimizes planning. On NAVSIM v1 it raises PDMS from 86.2 to 91.3, reaches 88.2 EPDMS on NAVSIM v2, and averages 38.8 HD-Score on HUGSIM.

  8. arXiv

    Building inspection stops at pixels: pulling Gaussian centers onto real surfaces cuts height error to 0.33%

    Three-dimensional Gaussian reconstruction for building inspection no longer aims only at photorealistic views: with region and point-wise depth supervision, Gaussian centers are pulled toward physical surfaces, cutting depth error by 58.1% and 86.8% on synthetic UAV and smartphone-captured data, with a mean relative height-measurement error of 0.33%.

  9. arXiv

    RDGSplat decodes render-dedicated geometry on a frozen 3D foundation model, lifting WM2.0 on RE10K from 20.918 to 24.266 dB while leaving its metric predictions unchanged

    RDGSplat introduces a two-pass framework that decodes a render-dedicated geometry from a frozen 3D foundation model by duplicating pretrained decoders and optimizing them under photometric supervision alone, together with a target-pose conditioned adapter that reformulates the representation those decoders read, while the model's metric predictions remain unchanged. Across four benchmarks and three feed-forward backbones, it improves novel view synthesis, raising WM2.0 on RE10K from 20.918 to 24.266 dB with 205.5 M added trainable parameters, and the depth and pose the same model predicts are bit-exact.

  10. arXiv

    VolS-GS solves volumetric subsurface scattering with a differentiable finite-volume solver on the Gaussian occupancy domain, improving relighting on three OLAT benchmarks

    VolS-GS turns the spatial support of a Gaussian scene into an occupancy field on a coarse grid and runs a differentiable finite-volume light-transport solve on it, with a small network predicting per-Gaussian scattering and absorption coefficients so light can propagate through the object's interior; across three OLAT benchmarks it consistently improves relighting quality on held-out lights and views over GS3, SSS-GS, RNG, and SSD-GS.