Microsoft Research

Robust prediction from corrupted latents, not diffusion objectives, unlocks superior visual tokenizers for diffusion models.

Few-step diffusion models can achieve state-of-the-art 4-NFE generation without auxiliary score models or adversarial training, matching or exceeding multi-step teachers.

Direct 2D spatial composition with four binding types unlocks unprecedented controllability in poster generation, outperforming prompt-based methods.

Ambiguous object selection in text-guided image editing is solved by directly annotating the canvas, achieving state-of-the-art performance in complex edits.

Discrete neural samplers achieve fixed-point convergence and drastically reduce mode collapse by leveraging Gibbs sampling within a masked diffusion framework.

Most video generation models fail to render text accurately, with even top performers exhibiting a 25% word error rate on a new benchmark designed for this critical oversight.

Append-only memory guarantees attention collapse in long-horizon video reasoning, but actively rewriting a bounded working memory from an external filesystem boosts hard-subset task accuracy by over 20%.

SELECT achieves a remarkable 2.8% boost in mIoU for class-incremental segmentation by intelligently transferring context from semantically related classes.

Fast-weight updates in LongVU-TTT enable MLLMs to retain crucial visual context, significantly boosting performance in long video understanding tasks.

Current video generation models struggle with visual reasoning, with the best achieving only 51% accuracy on a new benchmark designed to probe their capabilities.

Colorization methods can now adaptively handle both modern and historical grayscale images, reducing color artifacts and improving visual fidelity.

EndoCLIP can match expert endoscopists in classifying lesions, showcasing the transformative potential of linking clinical findings to images in colonoscopy.

Transitioning from 2D to 3D modeling reveals that fine details in monocular geometry can be captured with unprecedented fidelity.

ReViV reconstructs 4D viewer and view dynamics from a single monocular video, achieving unprecedented accuracy and speed without heavy task-specific priors.

Natural paper revisions can be harnessed to train AI agents for precise and context-aware editing of complex scientific diagrams.

HASTE enables rapid, accurate building damage assessment in disaster zones, delivering actionable insights within hours using minimal user input.

Current MLLMs miss critical insights from multi-view sports videos, but a new agentic framework boosts performance by over 14% through smarter view selection.

Temporal domain adaptation can dramatically enhance high-resolution climate projections, especially in challenging topographical regions.

Ink3D achieves a breakthrough in 3D asset creation, enabling the generation of complex textures that were previously unattainable with conventional methods.

Achieving up to 3.5% mAP gains and 1.3x higher throughput, RT-SFOD redefines the trade-offs in source-free object detection by being faster and more compact without sacrificing accuracy.

Bridging the Context Gap in T2I models, Qwen-Image-Agent achieves state-of-the-art performance by intelligently constructing context from user input and external sources.

Robots can now navigate crowded spaces more effectively by understanding human intentions, thanks to a new method that integrates rich visual cues into their decision-making process.

MambaRaw achieves a remarkable 1.4 dB increase in PSNR at low metadata bitrates while slashing coding latency by nearly 9%, setting a new benchmark in raw image reconstruction.

CoTriSyGen achieves unprecedented long-range coherence in video generation by integrating visual evidence into a dynamic memory system, drastically reducing identity drift across shots.

A million-scale dataset for identity-preserving video generation enables a new benchmark that outperforms existing models with minimal parameter overhead.

CapRL++ redefines caption quality through utility, enabling models to produce high-fidelity descriptions without the constraints of traditional supervised fine-tuning.

Skip the training and the hyperparameter tuning: Swift Sampling uses Taylor series to find the most informative frames in a video, beating existing methods with a 30x reduction in overhead.

Forget grid layouts: Map2World lets you generate consistent 3D worlds from arbitrary segment maps, offering unprecedented control and scalability.

Simple, artist-friendly quad meshes can now be automatically generated on 3D shapes using a diffusion model trained on a continuous surface representation, sidestepping the complexity of discrete mesh optimization.

Edit 3D assets with text prompts while actually preserving the original object's unchanged parts, thanks to a new masking strategy and training dataset.

Real-time, lightweight image compression is now possible with diffusion models, thanks to a novel architecture that swaps transformers for convolutions and prioritizes compression-focused pre-training.

Iterative visual refinement lets agents navigate dense coding IDEs with superhuman precision, outperforming single-shot methods and paving the way for more reliable software engineering agents.

Even when source data is protected, source-free domain adaptation leaks knowledge of source-exclusive classes into the target domain, creating a privacy risk that can be mitigated with adversarial unlearning.

Autonomous driving models can now achieve remarkable zero-shot generalization by leveraging the power of large-scale video generation models to jointly predict future actions and visuals.

GeoAI assistants remain unproductive because they lack a crucial agency layer for iterative human-AI collaboration, a gap this paper addresses with nine core primitives.

Synthetic motion data, when represented as optical flow, unlocks a new level of realism and control in video diffusion models, surpassing the limitations of real-world datasets.

Ditch mean pooling in your geospatial foundation models: richer pooling methods like GeM can boost accuracy by up to 5% and slash the geographic generalization gap by 40%.

LLMs can now more accurately answer questions on complex documents thanks to a new system that understands layout and hierarchical relationships between document components.

Achieve up to 57% better point cloud compression by combining the generalization of pretrained models with the robustness of implicit neural representations.

ImageNet-pretrained CNNs can spot looted archaeological sites from space with surprising accuracy, leaving traditional methods in the dust.