# Novel View Synthesis

Atlas was an autoregressive world model that generated frames one at a time, and users waited tens of seconds or even minutes for a video, so they planned camera paths cautiously. The new research preview Frameboy serves a highly optimized Atlas variant from cloud GPUs, streaming frames to the browser at up to 15 fps so the view updates continuously as the camera moves. The same real-time capability is integrated into the web app Viewfinder, where users build their own scenes frame by frame and then explore them live through a Frameboy overlay.

Atlas was an autoregressive world model that generated frames one at a time, and users waited tens of seconds or even minutes for a video, so they planned camera paths cautiously. The new research preview Frameboy serves a highly optimized Atlas variant from cloud GPUs, streaming frames to the browser at up to 15 fps so the view updates continuously as the camera moves. The same real-time capability is integrated into the web app Viewfinder, where users build their own scenes frame by frame and then explore them live through a Frameboy overlay.

D3WM splits next-frame prediction into a deterministic part and an uncertain part, using a learned uncertainty map to steer a diffusion model toward occlusions and dynamic textures, and it beats comparison world models on both robotic manipulation and camera motion.

D3WM splits next-frame prediction into a deterministic part and an uncertain part, using a learned uncertainty map to steer a diffusion model toward occlusions and dynamic textures, and it beats comparison world models on both robotic manipulation and camera motion.

The authors present OX-NeRF, which fuses a cross-scene convolutional encoder with a per-scene multi-resolution hash grid into an MLP and optimises them jointly by self-supervision, achieving higher 3D reconstruction accuracy than existing self-supervised radiance field methods on parallel-beam and cone-beam X-ray datasets in the ultra-sparse regime of four to ten projections.

The authors present OX-NeRF, which fuses a cross-scene convolutional encoder with a per-scene multi-resolution hash grid into an MLP and optimises them jointly by self-supervision, achieving higher 3D reconstruction accuracy than existing self-supervised radiance field methods on parallel-beam and cone-beam X-ray datasets in the ultra-sparse regime of four to ten projections.

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.

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.

MoonGS is a feed-forward 3D Gaussian Splatting framework for lunar scenes: given only two images it predicts pixel-aligned Gaussian primitives in a single forward pass with no per-scene optimization, using robust depth features from a frozen vision foundation model (DUSt3R, or VGGT), semantic priors plus a semantic ranking loss to regularize weakly textured background depth, and an entropy-guided heuristic resampling strategy to augment sparse observations. On the LuSNAR benchmark and the authors' weak-texture MoonBlender dataset it outperforms feed-forward NeRF/3DGS baselines on PSNR, SSIM and LPIPS while keeping sub-second inference.

MoonGS is a feed-forward 3D Gaussian Splatting framework for lunar scenes: given only two images it predicts pixel-aligned Gaussian primitives in a single forward pass with no per-scene optimization, using robust depth features from a frozen vision foundation model (DUSt3R, or VGGT), semantic priors plus a semantic ranking loss to regularize weakly textured background depth, and an entropy-guided heuristic resampling strategy to augment sparse observations. On the LuSNAR benchmark and the authors' weak-texture MoonBlender dataset it outperforms feed-forward NeRF/3DGS baselines on PSNR, SSIM and LPIPS while keeping sub-second inference.

The authors present SNAP, a self-supervised encoder-decoder transformer in which a pose-free encoder aggregates frozen DINOv3 patch features via full cross-view self-attention, while a PRoPE pose-conditioned decoder with a BlockNN mask reduces target-view self-attention to a pointwise residual MLP and is trained to reconstruct frozen DINOv3 latent features with a first-order spatial finite-difference penalty; after training on roughly 117,000 sequences from RealEstate10K, DL3DV, and Co3Dv2, the frozen encoder matches or exceeds self-supervised baselines on point correspondence, visual localization, relative pose, relative depth, and MimicGen robot manipulation, and degrades more gracefully than 2D semantic encoders under camera shifts.

The authors present SNAP, a self-supervised encoder-decoder transformer in which a pose-free encoder aggregates frozen DINOv3 patch features via full cross-view self-attention, while a PRoPE pose-conditioned decoder with a BlockNN mask reduces target-view self-attention to a pointwise residual MLP and is trained to reconstruct frozen DINOv3 latent features with a first-order spatial finite-difference penalty; after training on roughly 117,000 sequences from RealEstate10K, DL3DV, and Co3Dv2, the frozen encoder matches or exceeds self-supervised baselines on point correspondence, visual localization, relative pose, relative depth, and MimicGen robot manipulation, and degrades more gracefully than 2D semantic encoders under camera shifts.