Preprint of today: Chen et al., "Towards Physical Understanding in Video Generation: A 3D Point Regularization Approach" -- Let video models also diffuse pixel-aligned point clouds via augmentation and regularization --> videos that make more sense

Preprint of today: Shen et al., "Seeing World Dynamics in a Nutshell" -- Feed-forward Dynamic Gaussian Estimator for videos. Estimates pixel-aligned Gaussians, then deforms them to adhere to depth and flow from Foundational models.

Preprint of today: Chefer et al., "VideoJAM: Joint Appearance-Motion Representations for Enhanced Motion Generation in Video Models" -- Ensemble conditioning of text, motion, and both, and teaching a video model to both results in much more natural video.

Preprint of today: Govindarajan, Rebain et al., "Radiant Foam: Real-Time Differentiable Ray Tracing" -- You can do volumetric ray tracing with a learned representation using meshes (foams), even faster than rasterizing with Gaussian Splats.

Preprint of today: Guizilini et al., "Zero-Shot Novel View and Depth Synthesis with Multi-View Geometric Diffusion" -- Generalizable view-conditioned novel-view synthesis diffusion model. The results aren't perfect, but give some nice gains from SOTA.

Preprint of (not) today: Liang et al., "Feed-Forward Bullet-Time Reconstruction of Dynamic Scenes from Monocular Videos" -- A ViT that takes multiple frames (and their time andpose in plücker representation) outputs 3D Gaussians at queried time.

Preprint of today: Jäger et al., "FeatureGS: Eigenvalue-Feature Optim. in 3D Gaussian Splatting for Geom. Accurate and Artifact-Reduced Recon." -- Aligns Gaussian shapes with how neighbouring Gaussians are arranged via eigenvalues. Byebye floaters!

RT @ducha_aiki: While you are finishing CVPR rebuttal - consider submitting papers to Image Matching Workshop 2025 @CVPR #CVPR2025 Deadli…

Paper of (not) today: Chen et al., "On the Trajectory Regularity of ODE-based Diffusion Sampling" -- An interesting paper that highlights the regularity of diffusion trajectories. And how we can do better than simply going about this in equal steps.

@DavidSHolz I can't edit because I don't pay :( the proper link. Sorry!

Preprint 2/2 of today: Ren et al., "Improving Tropical Cyclone Forecasting With Video Diffusion Models" -- Video diffusion models can help predict cyclones! Note: baseline is also a diffusion model.

Preprint 1/2 of today: Elflein et al, "Light3R-SfM: Towards Feed-forward Structure-from-Motion" -- Feed-forward SfM using Dust3r style arch, but with latent-space alignment. Somewhat similar direction as Fast3R (yesterday's post)

Preprint of the day: Kinakh et al., "Binary Diffusion Probabilistic Model" -- Images are discrete and binary. Shouldn't we take that into account when we do probabilistic modeling? This work uses XOR-based noise transformations instead of Gaussians.

Paper of the day: Wang et al., "Continuous 3D Perception Model with Persistent State" -- RNN state model that can be "read out" as 3D pointmaps.

Preprint of the day: Raphaeli et al., "SILO: Solving Inverse Problems with Latent Operators" -- Learn an image degrader in the latent space for inverse problems for both speed and quality.

Preprint of the day: Wen et al., "FoundationStereo: Zero-Shot Stereo Matching" -- Extending DepthAnythingV2 to do stereo with a learned cost-volume filter. Cool visuals on the website!

Preprint of the day: Motamed et. al, "Physics IQ Benchmark: Do generative video models learn physical principles from watching videos?" -- A benchmark to test whether video models really learned physics -- spoiler: they didn't (yet)

repost with video since i forgot to put that up the first time :(