Lin et al., "SDiT: Semantic Region-Adaptive for Diffusion Transformers" Segment your latent into regions, choose which region(s) to denoise based on "complexity" heuristic, then update the rest using past estimates. Ie, each pixel has its own denoising schedule.

Wu et al., "Motion Attribution for Video Generation" From which data do video models learn different types of motion? Finding this, via backtracking gradients, allows data curations and fine-tuning of models to "better" motion.

Tong and Chang et al., "CoF-T2I: Video Models as Pure Visual Reasoners for Text-to-Image Generation" Curated dataset (with vlm etc) to use "frames" as thought chains for text-to-image generations.

Sucar and Insafutdinov et al., “V-DPM: 4D Video Reconstruction with Dynamic Point Maps” VGGT + time-conditioned point map estimates. Similar to Monst3r, but with VGGT. Trained to map to a canonical view, and static points live at a “canonical time”.

Zhao and Wei et al., "Spatia: Video Generation with Updatable Spatial Memory" "Memory" for video models via point cloud-conditioned video generation. I am obviously still biased towards having these "explicit" 3D stuff.

Wu et al., "From Rays to Projections: Better Inputs for Feed-Forward View Synthesis" Train a DiT to recover the full image from point cloud rasters. Is 3d "cueing" all we need? In a similar spirit to other works that "fix" rough 3D renders.

Wang et al., "MoE3D: A Mixture-of-Experts Module for 3D Reconstruction" Flying pixels in DPT-based models are coming from the fact that DPT modules are convolutional. Introducing MoEs allows you to circumvent that. So...sort of bilateral filtering?

Jiang et al., "ImLoc: Revisiting Visual Localization with Image-based Representation" Visual localization and image matching are always on my radar. Even with "modern" methods, perhaps we'd still want traditional image-based techniques well tied together.

Yu, Lin, and Wang et al., "InfiniDepth: Arbitrary-Resolution and Fine-Grained Depth Estimation with Neural Implicit Fields" Power of ViTs (DINOv3) + Neural Field decoder for resolution-free depth estimates.

Huang and Li et al., “How Much 3D Do Video Foundation Models Encode?” A lot. Which makes sense, given the way 3D foundational models are trained is VERY similar to what video models would see.

Tedla et al., "Generating the Past, Present and Future from a Motion-Blurred Image" Fine-tune a video model to create the video of what would've created the blurry image. So "live" photos from motion blur I guess? Neat :)

Brady et al., "Generation is Required for Data-Efficient Perception" Paper argues that compositional generalization is infeasible in a pure encoder setup, whereas with a decoder, it's easy. Not sure if infeasible, but certainly easier the other way around.

Alzugaray et al., "ACE-SLAM: Scene Coordinate Regression for Neural Implicit Real-Time SLAM" Scene Coordinate Regression networks have shown quite impressive performance when it comes to efficiency. Now, here's how you do SLAM with it. Not as accurate, but MUCH more lean.

Mazur et al., "4D Primitive-Mâché: Glueing Primitives for Persistent 4D Scene Reconstruction" Feed-forward reconstruction + segmentation-based "glueing", ie, primitives (segments) that move together according to correspondences share motion.

Segre and Hirschorn et al., "Multi-View Foundation Models" Train an adapter to 2D foundational models like DINO/SAM/CLIP that allows turning them into "multi-view" versions. More reliable results when doing multi-view tasks.

Metzger et al., "Elastic3D: Controllable Stereo Video Conversion with Guided Latent Decoding" Turn any video into a stereo video. Single-step diffusion + guided decoder, instead of explicit warping. The decoder does use epipolar geometry for guidance.

Po et al., "BAgger: Backwards Aggregation for Mitigating Drift in Autoregressive Video Diffusion Models" To mitigate drifting in autoregressive video gen, train with drifts, and also teach how to correct drifts via reversing videos.

Ramos et al., "Beyond the Noise: Aligning Prompts with Latent Representations in Diffusion Models" You can learn to measure image-text alignment early on in the denoising process. Allows early pruning of bad outcomes for Best-of-N generations.

Karhade et al., "Any4D: Unified Feed-Forward Metric 4D Reconstruction" Map Anything-style architecture for estimating everything geometry + scene flow. Impressive dense estimation results!

Zhou et al., "Scaling Zero-Shot Reference-to-Video Generation" Autogenerated mask-based training with masked frames serving as reference (with augmentations). Yes, I am posting this becuz I just liked the Einstein riding the bike video (and they promise inference release).

Meier et al., "LeAD-M3D: Leveraging Asymmetric Distillation for Real-time Monocular 3D Detection" Blend two images together (mixup) and make the student detect 3D objects in both images, while the teacher sees the clean ones. Surprised that this sort of blending (+ more) works!

Zeng et al., "NeuralRemaster: Phase-Preserving Diffusion for Structure-Aligned Generation" You can LoRA fine-tune existing models to only destroy the magnitude of the image while noising, which then leads to a structure-preserving image/video editor.

Xiao et al., "Video4Spatial: Towards Visuospatial Intelligence with Context-Guided Video Generation" Fine-tune a video model to generate videos based on context, solving object grounding and scene navigation. Another one in the trend of "let's generate solution videos"

Hu, Cheng, Yu et al., "VGGT4D: Mining Motion Cues in Visual Geometry Transformers for 4D Scene Reconstruction" Easi3r-style attention analysis and masking with mask refinement with VGGT. Also discards tokens related to dynamic points.

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.