Goldstein et al., "Flow Map Learning via Nongradient Vector Flow" Placing stop-gradient seems to be useful when training flow models. Here's now a theory on why.

Urbański and Maggiora et al., "Noise-Robust Conditional Flow Matching: Generating Clean Samples from Noisy Datasets" Data is often noisy. Naturally, you can fold your sensory noise into your flow model. This often still models noise, and you can stop denoising early to mitigate

Li et al., "JigShape: Evaluating Visual-Geometric Reasoning in VLMs through Jigsaw Puzzles" How well can VLMs do puzzles? -- apparently not that well. They are barely better than random draws for 8x8 sizes and most of them even at 4x4.

Pataki et al., "VidMap: Exploiting Temporal Structure for Video-Based Structure-from-Motion" SLAM uses temporal continuity but can drift; SfM can be out-of-order but ignores temporal continuity. In hindsight, we should obviously do both?

Geirhos, Li, Wiedemer, et al., "Visual prompt engineering for video models" When repurposing video models are reasoning engine, you can edit your input prompt so that it's "friendlier" to video models.

Lisowski and Smoliński et al., "TOM-GS: Editable Video Representation via Temporal Opacity Modulation of Static 3D Gaussians" Instead of having movements, you can also represent videos with just pure opacity changes. Makes me wonder if Gaussians ever have to move

Chen, Chen, Zhang, et al., "Engine-Native Editable 3D World Reconstruction with Objects and Lighting" Yep, extract point clouds, boxes, light, etc, as much as you can and then make GPT code it up in blender. Why not?

Weijler et al., "Latent Riemannian Flow Matching for Geometry-Grounded 3D Foundation Models" Denoising on the latent space of VGGT brings its 3D priors into play. To do so, Riemannian flow matching is used to stay on the manifold.

AlayaWorld Team, "AlayaWorld: Long-Horizon and Playable Video World Generation" Interactive 24 FPS* 720p model with 15B parameters. Open weights. Not bad for an open model. Uses depth anything and point clouds as spatial memory. BTW, "warp" not "wrap" -- this bothers me so much!

Xue et al., "DepthART: Scaling Foundation Monocular Depth to Tiny Models" Well-strategized distillation with camera-conditioned fine-tuning. 1000 FPS on RTX A6000, 200 FPS on Jetson Orin NX.

Gong et al., "MAGiSt3R: Multi-Agent Feed-forward 3D Reconstruction from Monocular RGB Videos" Use multiple agents to do feed-forward SLAM!

Elflein et al., "VGG-TTT: Offline Feed-Forward 3D Reconstruction at Scale" VGGT, but with Test-time training to compress KV space. I finally understood TTT today thanks to a collaborator. Very much reminds me of scene coordinate regression networks.

Deng, Li, Qiu, et al., "Glob3R: Global Structure-from-Motion with 3D Foundation Models" Another one showing that you SHOULD do global optimization and bundle adjustment (BA) with Feed-forward geometry estimators. Traditional formula: keyframes + view graphs + tracks + BA.

Reijalt et al., "On the Real-World Generalisability of Optical Flow Models" Guess what, optical flow datasets have saturated. And yes, if you are using RAFT, you are not missing much by not using more recent ones.

Ziliotto et al., "What VGGT Knows About Overlap: Probing Geometric Foundation Models for Co-Visibility" Feed-forward 3D models like VGGT show layer-wise emergent behaviors. You can treat layers as experts and train an MoE to do co-visibilty prediction, and outperform humans.

Song and Bonilla et al., "Track2Map: Online Deformable SLAM with Motion-Aware Pose Optimization in Robotic Surgery" CoTracker3 + Dynamic 3D Gaussians + Gating of camera pose based on flow stats. Nice non-rigid SLAM for endoscopy.

To be presented at ECCV'26 -- now with user studies (turns out ours is preferred by 90%) and more analysis. We were unsure before how to explain optimizing in Fourier space, but turns out it's plain-old preconditioning.

Kwang Moo Yi@kmyid.bsky.social · 8mo ago

Baek et al., "SONIC: Spectral Optimization of Noise for Inpainting with Consistency" Initial seed noise matters. And you can optimize it **without** any backprop through your denoiser via good-ol linearization. Importantly, you need to do this in the Fourier space.

Yuan et al., "StereoGS: Sparse-View 3D Gaussian Splatting via Stereo Priors" Cute idea -- stereo networks work well, why not embed them as regularizers when training 3DGS?

Zhang and Taubner et al., "ProxyPose: 6-DoF Pose Tracking via Video‑to‑Video Translation" Finetune (LoRA) a video generator to generate proxy cube videos demonstrating 6 degrees of freedom (DoF) to do point tracking in 3D. Impressive generalization.

Sinitsyn et al., "RayTun3R: Online Camera Adaptation in 3D Foundation Models" You can quickly tune Positional Encoding adapters (LoRA) to turn existing feed-forward geometry estimators for cameras other than pinhole ones, with correspondences etc with only the first few frames.

D’Urso et al., "Boosting 3D Foundation Models with Edge-based Pose Optimization" Instead of costly BA, minimize bi-direction edge alignment of the two images using your feed-forward geometry model.

Hirschorn et al., "SpheRoPE: Zero-Shot Optimization-Free 360◦ Panorama Generation with Spherical RoPE" Periodic RoPE + additional guidance by CFG using a prompt that encourages 360 panoramas. Allows training-free & optimization-free repurposing to generate 360 panoramas.

Qian et al., "Semantic-Aware, Physics-Informed, Geometry-Grounded Weather Video Synthesis" GPT4.1 + Simulated Gaussian Particles + Depth Anything v3 + Flux Kontext + Wan 2.1-VACE = some nice cool weather edits in videos. Composed of mostly existing methods, but tightly knit.

Barath "Rolling Shutter Relative Pose Estimation Made Practical" You often assume all pixels come at the same time, but that's actually not true, and it does matter for accurate camera localization, unfortunately. And now, proper math for a minimal solver to use with RANSAC.

Morbitzer and Evers et. al., "Future Dynamic 3D Reconstruction: A 3D World Model with Disentangled Ego-Motion" Interesting that the leap in feed-forward 3D came from predicting geometry first, then inferring cameras. Now the trend seems to go back to disentangling them.

Gilo et al., "FlowBender: Feedback-Aware Training for Self-Correcting Conditional Flows" Introduce cycle consistency as a feedback input during both training and inference for improved conditional flow models.

Xiao et al., "Triangular Consistency as a Universal Constraint for Learning Optical Flow" You can chain optical flow in various ways to form constraints. Yes, this seems very familiar, yet somehow we seem to have forgotten to do this VERY effective thing. It also requires care.

Dufour et al., "The FID Lottery: Quantifying Hidden Randomness in Generative Model Evaluation" We all know it's expensive to train multiple times, but we are now at a point where it is inevitable. Statistical significance should not be ignored. Don't bold over 1~2% differences.

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.