Zhang et al., "SparseSplat: Towards Applicable Feed-Forward 3D Gaussian Splatting with Pixel-Unaligned Prediction" Feed-forward 3D Gaussian Splatting is often pixel-aligned. Here's a non-aligned one, based on 3D point clouds and anchors.

Khangaonkar et al., "Multimodal Large Language Models Cannot Spot Spatial Inconsistencies" A benchmark for spotting spatial inconsistencies. MLLMs are still quite far from being accurate. Reminds me of various automated benchmarks that have recently been shown to be misleading.

Zhang et al., "Watch Before You Answer: Learning from Visually Grounded Post-Training" When post-training VQA, sometimes your answers can be inferred easily **without** the images. These shortcuts are harmful. Removing them can improve performance.

Ma et al., "SimpleProc: Fully Procedural Synthetic Data from Simple Rules for Multi-View Stereo" Turns out you can achieve state of the art, or even outperform it, with ZERO real data for multi-view stereo. Which actually does make sense since you are learning to solve geometry.

Huang et al., "UniRecGen: Unifying Multi-View 3D Reconstruction and Generation" Estimate 3D point clouds (maps) in both camera and object coordinates, which leads to better multi-view feed-forward reconstruction.

Newman et al., "Video Models Reason Early: Exploiting Plan Commitment for Maze Solving" Similar to the chain of steps paper, the so-called "reasoning" of video models seems to happen early, allowing exploitation of early steps for better problem solving.

Li and Luo et al., "Benchmarking PhD-Level Coding in 3D Geometric Computer Vision" Am I replaced yet? Not yet! 3D seems still hard. I guess we still have to provide them with tools for now. But for how much longer?

Huang et al., "From None to All: Self-Supervised 3D Reconstruction via Novel View Synthesis" Learn to predict Gaussians, cameras, and depth from collections of images. Without any priors, from scratch, just geometric constraints in its architecture.

An et al., "VGGRPO: Towards World-Consistent Video Generation with 4D Latent Reward" Connect (read out) video model latents to Any4D, then perform geometric RL feedback. Enhances geometric consistency in videos.

Zhu et al., "GaussFusion: Improving 3D Reconstruction in the Wild with Geometry-Informed Video Generator" Another diffusion "fixer" BUT with now a geometry buffer (providing all the info about what IS being rendered)

Qiao et al., "AdvSplat: Adversarial Attacks on Feed-Forward Gaussian Splatting Models" Feed-forward geometric models are great, but they can easily suffer from adversarial attacks, just like any neural network-based regressors.

Guruprasad et al., "CAM3R: Camera-Agnostic Model for 3D Reconstruction" Estimate camera as well, encoded as *spherical harmonics*, which allows various cameras to be dealt with a single model.

Hu and Han, "SGAD-SLAM: Splatting Gaussians at Adjusted Depth for Better Radiance Fields in RGBD SLAM" Generalized ICP + Depth-wise adjusted pixel-aligned Gaussians. You don't need all the fancy bells and whistles, just a "tight" implementation.

Zhang and Liang et al., "2K Retrofit: Entropy-Guided Efficient Sparse Refinement for High-Resolution 3D Geometry Prediction" Use foundational models, upsample, and "fix" the high-entropy regions (which coincide well with estimation errors) via 3D Sparse Convs.

Li and Zhu et al., "DROID-SLAM in the Wild" DROID-SLAM, but now with monocular depth estimators and uncertainty estimates, so you can run on any RGB video.

Yu et al., "MosaicMem: Hybrid Spatial Memory for Controllable Video World Models" A patch-based spatial memory that you raster into views + glues to make things work.

Wang et al., "Demystifying Video Reasoning" A deep dive into how video models "reason" to generate solution videos -- they seem to be doing "chain of steps" thinking, where denoising steps behave as if they are chain of thoughts.

Chen et al., "GGPT: Geometry-Grounded Point Transformer" Feed-forward models -> RoMA -> Sparse BA for accurate SfM, then refine it via point transformer. An extensive ablation study backing up design choices.

Sun et al., "VGGT-World: Transforming VGGT into an Autoregressive Geometry World Model" Train an autoregressive DiT in the VGGT latent space, with epsilon prediction, i.e., if you want geometric forecasting, use a geometric foundational model.

Aydemir et al., "Real-World Point Tracking with Verifier-Guided Pseudo-Labeling" Train a verifier with multiple point tracking methods. Use the verifier to ensemble them to create pseudo labels to train your point tracker now with more data.

Xiao et al., "Spectral Probing of Feature Upsamplers in 2D-to-3D Scene Reconstruction" A deep analysis of how upsamplers perform. Conclusion: learned ones, at least for now, show a limited advantage when tied properly with the reconstructor.

Cheng et al., "ReCoSplat: Autoregressive Feed-Forward Gaussian Splatting via Render-and-Compare" Feed-forward Gaussian Splats have errors. Render current estimates to incoming views, and use those together for better reconstruction. Ie, render and compare.

Mukhopadhyay* and Udhayanan*, et al., "Scale Space Diffusion" Scale space + Diffusion. Scale space is back. "Destroy" information by not just noising, but also scaling.

Lu et al., "OVGGT: O(1) Constant-Cost Streaming Visual Geometry Transformer" Updating KV caches with the most important tokens, using a fixed budget, with some anchors to prevent drifting. Constant budget, improved speed & quality.

Cheng and Shaikh et al., "MERG3R: A Divide-and-Conquer Approach to Large-Scale Neural Visual Geometry" Simple, practical idea that works. Sort->split->merge for faster, scalable reconstruction.

Jain et al., "NERFIFY: A Multi-Agent Framework for Turning NeRF Papers into Code" Given what we have now, it makes sense that we should no longer spend time reproducing work. Let agents read paper, code, verify with nerfstudio.

Zhu et al., "Marginalized Bundle Adjustment: Multi-View Camera Pose from Monocular Depth Estimates" With monocular depth estimators being more and more accurate, they are often better for SfM in many cases. Good, but not the best for IMC phototourism yet!

Cong et al., "Flow3r: Factored Flow Prediction for Scalable Visual Geometry Learning" More data is better, and you can use unlabeled videos with off-the-shelf dense flow predictors to do this. 800k unlabeled video to further enhance geometry estimation.

Adamkiewicz et al., "When Pretty Isn't Useful: Investigating Why Modern Text-to-Image Models Fail as Reliable Training Data Generators" Interesting that while we have "better" image generators, their usefulness as synthetic data generators is declining. Do we need a pivot?

Xie, Sun, Neall et al., "Generated Reality: Human-centric World Simulation using Interactive Video Generation with Hand and Camera Control" Camera & 3D hand pose conditioned video generation with DiTs.

Ye et al., "World Action Models are Zero-shot Policies" Video + Action training. A LOT of engineering gems that allow making it work with an actual robot.

Liu at al., "Unpaired Image-to-Image Translation via a Self-Supervised Semantic Bridge" Train a diffusion bridge to go to Self-Supervised foundational features and back. Allows image-to-image translation.

Ke et al., "CAPA: Depth Completion as Parameter-Efficient Test-Time Adaptation" Fine-tune your foundational model at test time with sparse measurements. Makes a lot of sense if you have, e.g. Lidar measurements with you.

Yue et al., "Image Generation with a Sphere Encoder" Sampling Gaussian noise independently for each pixel causes you to actually sample more-or-less on a sphere. So let our latent space be on a sphere.

Qin and Sun et al., "Variation-aware Flexible 3D Gaussian Editing" Distill 2D edits into a feed-forward model that predicts edit fields. Quick 3D editing of 3D Gaussians.

Bajpai et al., "FastFlow: Accelerating The Generative Flow Matching Models with Bandit Inference" You can use Multi-Armed Bandits with Flow Matching models to accelerate inference by 2.6x, all without any training and minimal overhead.

Luo et al., "4RC: 4D Reconstruction via Conditional Querying Anytime and Anywhere" Encode videos into 4D latents, then query motion and geometry between any two frames.

Luo et al., "4RC: 4D Reconstruction via Conditional Querying Anytime and Anywhere" Encode into a 4D latents, which is used to conditionally decode geometry and motion.

Mauel and Hübers et al., "Foundation Inference Models for Ordinary Differential Equations" Feed-forward solver for generic ODEs. Also easily fine-tunable for a given task. These things always fascinate me. Sort of learning to learn (solve?)

Shavin and Benaim, "Splat and Distill: Augmenting Teachers with Feed-Forward 3D Reconstruction For 3D-Aware Distillation" When distilling vision foundation models with a focus on geometric consistency, insert a feed-forward Gaussian Splatting in the middle.

Antsfeld et al., "S-MUSt3R: Sliding Multi-view 3D Reconstruction" Sliding window strategy for long sequences. Makes a lot of sense for practical applications -- uses 60 frames at a time with 30 frames overlap, with light loop closure.

Zhao et al., "Sparsely Supervised Diffusion" Simple, but seemingly effective idea. Just randomly masking your diffusion supervision seems to lead to less overfitting (of course?). Not to be confused with masked diffusion, this is simply during training.

Hsiao et al., "3DGS^2 -TR: Scalable Second-Order Trust-Region Method for 3D Gaussian Splatting" Who else likes a nice optimization paper? Gaussian Splatting optimizer that approximates curvature using only the diagonal of the Hessian matrix, efficiently via Hutchinson's method.

Du, Ye, Cong et al., "VideoGPA: Distilling Geometry Priors for 3D-Consistent Video Generation" Video models still suffer 3D inconsistencies. Generate video -> VGGT -> DPO for better 3D consistency. My personal question is, will it ever get perfect?

Lu et al., "One-step Latent-free Image Generation with Pixel Mean Flows" Mean flows, but now in pixel space. Single-step generation with raw pixels has come a long way ;)

Zhou et al., "FreeFix: Boosting 3D Gaussian Splatting via Fine-Tuning-Free Diffusion Models" Training-free method to "fix" 3DGS with diffusion. Render novel views --> SD Edit+Guidance --> refine.

Maggio and Carlone, "VGGT-SLAM 2.0: Real time Dense Feed-forward Scene Reconstruction" Improved version of VGGT-based SLAM. What I find really interesting is Layer 22 -- it shows correspondences and can be used to test for overlaps!

Tedla et al., "Learning to Refocus with Video Diffusion Models" Lots of videos have moments where the camera"refocuses". Natural that video models can be used to refocus images ;)

Mahapatra et al., "DreamLoop: Controllable Cinemagraph Generation from a Single Photograph" Layout and trajectory-controlled video generator to create video loops from a single image. Neat application, well-engineered.

Dai et al., "Keyframe-Based Feed-Forward Visual Odometry" Keyframes have been a critical idea for SLAM. How you extract and use them still matters in the era of VGGT. A Reinforcement Learning attempt at it.