Rädsch et al., "Physics-IQ Verified" More accurate prompts, better metrics, and artifacts removed. Provides a slightly different story on which models are better.

Hwang et al., "NoiseTilt: Noise-Tilted Reverse Kernels for Diffusion Reward Alignment" Guiding diffusion models can often lead to off-manifold samples. This one proposes to prevent moves from escaping the expected noise stats via clipping outliers.

Burns and Fridovich-Keil, "When, why, and how do diffusion posterior samplers fail? A finite-sample lens" A nice read on where various (approximate) denoising formulations fall short, and a look into how bad they are. Might be amortizing in large datasets, but still nice to know

Villeroché et al., "Zero-shot generalization of transformer neural operators to larger domains" Rotary Positional Encoding (RoPE) works well, but is non-local, which can make extrapolation hard. You can add locality via a position bias.

Zang et al., "World Tracing: Generative Pixel-Aligned Geometry Beyond the Visible" A Diffusion Transformer that estimates multiple layers of depth to further estimate occluded parts as well.

Zeng et al., “i1: A Simple and Fully Open Recipe for Strong Text-to-Image Models” A fully reproducible recipe & code & weights and everything for a truly open text-to-image model. A LOT of interesting findings.

Dabhi and Gill et al., "2D-LFM: Lifting Foundation Model without 3D Supervision" Simply using transformers to do 2D-to-3D lifting of 2D landmarks fails by construction due to the permutation equivariance of the architecture -- inject positional encoding in multiple layers to fix

Xiangli et al., "Honey, I Shrunk the Arc de Triomphe!" Metric depth estimators aren't actually metric. With curated, scaled data, they can be adapted to be better.

Esmati and Nath et al., "The Invisible Hand of Physics: When Video Diffusion Models Know More Than They Show" You can use inversion to retrieve feature representations for a video, which can be linearly decoded into physical plausibility -- if you use enough steps not shortcuts

Veicht et al., "ZipSplat: Fewer Gaussians, Better Splats" Non pixel-aligned feed forward Gaussian Splatting. With K-means to cluster DiT tokens to compress at arbitrary ratio, decoded into 3D Gaussians.

Kim and Cheng and Kong, et al., "MLP Splatting: Object-Centric Neural Fields" Instead of Gaussians, use an MLP, ie, learn your kernels. One thing I want to mention -- while the title mentions object-centric, not sure if it truly is. To me, this is more like a learned kernel.

He and Luo et al., "Thinking in Blender: Staged Executable Inverse Graphics with Vision-Language Models" Inverse graphics into a blender model via VLM, with multiple staged goals. Not there yet, but also not entirely off.

Burzio and Fischer et al., "Deja View: Looping Transformers for Multi-View 3D Reconstruction" Recurrent execution of the same small transformer to do feed-forward point map (ray + depth) estimation. 3D inference requires iterative solvers. Is this one also a fixed-point solver?

Pan et al., "Global Structure-from-Motion Meets Feedforward Reconstruction" Pi3 + Global optimization, but a different flavor than the one I posted yesterday. So feed-forward + global optimization now reliably beats COLMAP.

Tse et al., "Learning Scene Coordinate Reconstruction from Unposed Images via Pose Graph Optimization" Extend Scene Coordinate Regression networks with Pose Graph Optimization (global optimization). Poses that give comparable reconstruction performance to COLMAP but faster.

Xu et al., "R3: 3D Reconstruction via Relative Regression" Estimate relative pose and depth maps, keyframe via kv caching + confidence estimation -- more accurate estimates, in real time.

Taher et al., "KV-Tracker: Real-Time Pose Tracking with Transformers" Instead of having a separate streaming state, you can KV-cache to get real-time tracking with feed-forward models like pi3 and VGGT. (I missed this work last Dec, but luckily found it in the CVPR proceedings)

Zheng et al., "Good Token Hunting: A Hitchhiker's Guide to Token Selection for Visual Geometry Transformers" Choose diverse frames, downsample differently according to depth -- sometimes even leads to slight performance gains.

Kocur, Aung, and Flood, et al., "Depth2Pose: A Pose-Based Benchmark for Monocular Depth Estimation without Ground-Truth Depth" You can use camera poses to benchmark depth estimators as well. Each model seem ot have their own bias towards which kind of scene they are best at.

Paul and Kaushik et al., “Can These Views Be One Scene? Evaluating Multiview 3D Consistency when 3D Foundation Models Hallucinate” Automated 3D consistency metrics based on “modern” models do NOT align with human evaluation, but COLMAP-based ones do. You can never escape COLMAP

Pearl and Esposito et al., "Learn2Splat: Extending the Horizon of Learned 3DGS Optimization" A learned optimizer for 3D Gaussian Splatting. Allows faster early convergence.

Wang et al., “VGGT-Ω” Scaling up VGGT with more data, unlabelled videos, self-supervision, VLM-based data curation, and architectural simplification/enhancements. vggt-omega.github.io

Nam et al., "TrackCraft3R: Repurposing Video Diffusion Transformers for Dense 3D Tracking" Lora-finetune a video model to take in dense point map estimates of each frame, with proper alignment and encoding == significant 3D tracking improvement.

Curvo et al., "Follow the Mean: Reference-Guided Flow Matching" Quoting exacts words from the paper -- "Guide with examples, not rewards"

Tang and Fu et al., "Spark3R: Asymmetric Token Reduction Makes Fast Feed-Forward 3D Reconstruction" Simple interesting finding -- you can compress key-value tokens aggressively for various feed-forward geometry models (VGGT, pi3, DepthAnything3). Architectural redundancy?

Ren et al., "Large Language Models are Universal Reasoners for Visual Generation" (M)LLMs can easily reason about prompt adherence of generation outcomes. Make a feedback loop to correct itself.

Tran et al., "Robust Fundamental Matrix Estimation from Single Image Motion Blur" With everything generative models and learned, this one is a nice, refreshing traditional topic -- extracting fundamental matrices of your camera motion just from how the pixels smeared.

Cohen et al., "Colorful-Noise: Training-Free Low-Frequency Noise Manipulation for Color-Based Conditional Image Generation" Encode your desired reference image into latent space, use only its low-frequency components, and fill the rest with white noise. --> Preserve structure.

Lin and Liu et al., "Vista4DVideo Reshooting with 4D Point Clouds" Feed-forward 4D reconstruction + masking and rasterization into a novel view, then use a video generator conditioned on this novel view render, masks, and the original video for novel view rendering.

Li et al., "Exploring Time Conditioning in Diffusion Generative Models from Disjoint Diffusion Data Manifolds" Time conditioning, and the way we architecturally enforce it, has been a core ingredient in making denoising models work. But perhaps that was a workaround, not a must.

Li et al., "Long-Tail Internet Photo Reconstruction" MegaDepth-X, a new curated dataset for in-the-wild reconstruction that enhances long-tailed distributions. Many feed-forward estimators are colmap distillations, and this one makes the data prep more modern, which matters.

Gabeur, Long, and Peng, et al., "Vision Banana Image Generators are Generalist Vision Learners" Instruction-tune nano banana pro to get SOTA on various tasks. Unfortunately, the report lacks details to not just reproduce, but also to understand the scientific merits.

Chen et al., "AnyRecon: Arbitrary-View 3D Reconstruction with Video Diffusion Model" Another "rasterize & fix" method. In addition to the common point map rasterization, add relevant views as anchors in your conditioning signal for conditional video generation.

Guan et al., "Latent-Compressed Variational Autoencoder for Video Diffusion Models" Video latents show unstructured frequency decomposition, indicating redundancy --> can be removed via wavelet transforms.

Briq et al., “The Amazing Stability of Flow Matching” The attached image explains it all (with minor caption error though) -- training flow matching models with less data, “different” data, disjoint subsets, different architectures all lead to similar results.

Ren and Tyszkiewicz, et al., "TokenGS: Decoupling 3D Gaussian Prediction from Pixels with Learnable Tokens" DETR-style feed-forward 3D GS, with test-time token adaptation.

Shen and Ren et al., "Lyra 2.0: Explorable Generative 3D Worlds" Explicit 3D memory, overlap-based frame extraction for conditioning with dense warping, using 3D Foundational models. I guess we still have need for explicit 3D.

Tran and Košecká, "PointSplat: Efficient Geometry-Driven Pruning and Transformer Refinement for 3D Gaussian Splatting" Prune, but then refine Gaussians in 3D with a point transformer.

Shabanov et al., "Free-Range Gaussians: Non-Grid-Aligned Generative 3D Gaussian Reconstruction" Denosing 3D Gaussians with DiT conditioned on multiview images for 3D assets -- via hierarchical Level of Detail representation for 3D Gaussians.

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.