Yang et al., "Weight Space Representation Learning with Neural Fields" A multiplicative Low Rank Adaptation (LoRA) strategy that works well for denoising on the network weight space. You can now denoise neural fields ;)

Wang and Agapito, "AMB3R: Accurate Feed-forward Metric-scale 3D Reconstruction with Backend" Beyond feed-forward 3D geometry estimators, this one also has a "backend" transformer that keeps a serialized 3D representation to close the loop.

Park et al., "Deep Parameter Interpolation for Scalar Conditioning" Here's a neat simple and effective way to condition models based on scalar inputs -- "interpolate" between deep network parameters! Makes a lot of sense given how MoEs work so well.

Meng et al., "Training-Free Generation of Diverse and High-Fidelity Images via Prompt Semantic Space Optimization" Guiding generation can enhance diversity, but it often comes at the cost of quality, or it often does not work. You can actually "optimize prompts" to avoid that.

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.

Chambon et al., "NAF: Zero-Shot Feature Upsampling via Neighborhood Attention Filtering" Essentially learning create an upsample kernel via looking at encoded images. Generalizes well to various foundational models. + it beats bicubic upsampling.

Maruani et al., "Illustrator’s Depth: Monocular Layer Index Prediction for Image Decomposition" Fine-tune a depth estimator to SVG layers for the ordinal depth estimator. Very cute outcomes with highly useful results.

SAM 3D Team, "SAM 3D: 3Dfy Anything in Images" image (point map) + mask -> Transformer -> pose / voxel -> transformer (image/mask/voxel) -> mesh / splat. Staged training with synthetic and real data & RL. I wish I could see more failure examples to know the limits.

Edstedt el al., "RoMa v2: Harder Better Faster Denser Feature Matching" V2 of RoMA is out :) 1.7x faster with improved performance! Dense feature matching, still outperforms feed-forward 3D methods by a far in outdoor scenarios.

Chen et al., "Co-Me: Confidence-Guided Token Merging for Visual Geometric Transformers" Train a confidence predictor for tokens and merge low-confidence ones for acceleration -> faster reconstruction with VGGT/MapAnything.

Vecchio et al, "Φeat: Physically-Grounded Feature Representation" Foundational backbone, finetuned DINOv3, trained with synthetic renders of materials, EMA student-teacher training with multiple losses.

Wu et al., "Equivariant Sampling for Improving Diffusion Model-based Image Restoration" Very simple idea -- for inverse problems, you can introduce equivariant augmentations which do not change the solution, and find a solution that matches everything. Augmentation in denoising.

Lin and Chen and Liew and Chen, et al., and Kang "Depth Anything 3: Recovering the Visual Space from Any Views" VGGT-like architecture, but simplified to estimate depth and ray maps (not point maps). Uses teacher-student training of Depth Anything v2.

Singer and Rotstein et al., "Time-to-Move: Training-Free Motion Controlled Video Generation via Dual-Clock Denoising" Make a rough warp, push it through Image-to-Video model with denoise together up until a timestep, then let it finish the rest without interference.

Yang et al., "UltraGS: Gaussian Splatting for Ultrasound Novel View Synthesis" 3D Gaussian Splatting for Ultrasounds. Tailored parameterization and Field-of-view adaptation for Ultrasound use cases.

Gutflaish and Kachlon et al., "Generating an Image From 1,000 Words: Enhancing Text-to-Image With Structured Captions" Is all you need just very good captioning? Makes a lot of sense given that this makes your problem closer to a 1-to-1 mapping instead of many-to-1.

Wang et al., "Seeds of Structure: Patch PCA Reveals Universal Compositional Cues in Diffusion Models" The existence of single (few) step denoisers, and many recent works hinted at this, but another one. You can decode the image structure from the initial noise fairly easily.

Ren and Wen et al., "FastGS: Training 3D Gaussian Splatting in 100 Seconds" I like simple ideas -- this one says you should consider multiple views when you prune/clone, which allows fewer Gaussians to be used for training.

Nguyen et al., "IBGS: Image-Based Gaussian Splatting" Gaussian Splatting + Image-based rendering (ie, copy things over directly from nearby views). When your Gaussians cannot describe highlights, let your nearby images guide you.

Gao and Mao et al., "Seeing the Wind from a Falling Leaf" Extract Dynamic 3D Gaussians for an object -> Vision Language Models to extract physics parameters -> model force field (wind). Leads to some fun.

Zhou et al., "PAGE-4D: Disentangled Pose and Geometry Estimation for 4D Perception" VGGT extended to dynamic scenes with a dynamic mask predictor.

Pfrommer et al., "Is Your Diffusion Model Actually Denoising?" Apparently no. This indeed matches the empirical behavior of these models that I experienced. These models are approximate, but then what actually is their mathematical property?

Tesfaldet et al., "Generative Point Tracking with Flow Matching" Tracking, waaaaaay back in the days, used to be solved using sampling methods. They are now back. Also reminds me of my first major conference work, where I looked into how much impact the initial target point has.

Stary and Gaubil et al., "Understanding multi-view transformers" We use Dust3r as a black box. This work looks under the hood at what is going on. The internal representations seem to "iteratively" refine towards the final answer. Quite similar to what goes on in point cloud net

Goren and Yehezkel et al., "Visual Diffusion Models are Geometric Solvers" Note: this paper does not claim that diffusion models are better; in fact, specialized models are. Just shows the potential that you can use diffusion models to solve geometric problems.

Luo et al., "Self-diffusion for Solving Inverse Problems" Pretty much a deep image prior for denoising models. Without ANY data, with a single image, you can train a denoiser via diffusion training, and it just magically learns to solve inverse problems.

Bai et al., "Positional Encoding Field" Make your RoPE encoding 3D by including a z axis, then manipulate your image by simply manipulating your positional encoding in 3D --> novel view synthesis. Neat idea.

Mao et al., "PoseCrafter: Extreme Pose Estimation with Hybrid Video Synthesis" While not perfect, video models do an okay job of creating novel views. Use them to "bridge" between extreme views for pose estimation.

Choudhury and Kim et al., "Accelerating Vision Transformers With Adaptive Patch Sizes" Transformer patches don't need to be of uniform size -- choose sizes based on entropy --> faster training/inference. Are scale-spaces gonna make a comeback?

Riise et al., "Visual Autoregressive Models Beat Diffusion Models on Inference Time Scaling" Beam search with Autoregressive image generators with verifiers.

Hakie and Lu et al., "Fix False Transparency by Noise Guided Splatting" Pretty cute idea -- Gaussian splats are often transparent, although we don't want them to be. So, just fill your splats in with noise during optimization to make them non-transparent.

Alzayer et al., "Coupled Diffusion Sampling for Training-Free Multi-View Image Editing" You can "guide" diffusion models with different purposes by "coupling them". Our group did simply weighted averaging without math in vivid-123, but this is much more sound!

Bruns et al., "ACE-G: Improving Generalization of Scene Coordinate Regression Through Query Pre-Training" Train a scene coordinate regressor with "map codes" (ie, trainable inputs) so that you can train one generalizable regressor. Then, find these "map codes" to localize.

Shrivastava and Mehta et al., "Point Prompting: Counterfactual Tracking with Video Diffusion Models" Put a red dot where you want to track, and SDEdit the video with a video model --> zero-shot point tracking. Not as good as supervised ones, but zero-shot!

Yuan et al., "LikePhys: Evaluating intuitive physics understanding in video diffusion models via likelihood preference" I will keep promoting physics benchmark papers for video models until people stop claiming world models :) tl;dr -- Still not there yet.

Xu et al., "ReSplat: Learning Recurrent Gaussian Splats" Feed-forward Gaussian Splatting + Learned Corrector = Fast high-quality reconstruction. Uses global + kNN attention. Reminds me of pointnet++

Xu and Lin et al., "Pixel-Perfect Depth with Semantics-Prompted Diffusion Transformers" Append foundational features at the later stages when doing marigold-like denoising to get monocular depth. Simple straightforward idea that works.

Bamberger and Jones et al., "Carré du champ flow matching: better quality-generalisation tradeoff in generative models" Geometric regularization of the flow manifold. Boils down to adding anisotropic Gaussian Noise to flow matching training. Neat idea, enhances generalization.

Yugay and Nguyen et al., “Visual Odometry with Transformers” Instead of point maps, you can also directly output poses. This used to be much less accurate, but now it's the opposite. Simple architecture that directly predicts camera embeddings, which then regress rot and trans.

Chen et al., "TTT3R: 3D Reconstruction as Test-Time Training" Cut3R + gated updates for states (test-time training layers) = fast/efficient performance of cut3r, but with high-quality estimates.

Two today: Kim et al., "How Diffusion Models Memorize" and Song and Kim et al., "Selective Underfitting in Diffusion Models" A deep dive into how memorization and generalization happen in diffusion models. Still trying to digest what these mean. Though-provoking.

Barroso-Laguna et al., "A Scene is Worth a Thousand Features: Feed-Forward Camera Localization from a Collection of Image Features" When contexting your feed-forward 3D point-map estimator, don't use full image pairs -- just randomly subsample! -> fast compute, more images.