🏢 Zhejiang University

Model Merging: An unlearning system, which combines specialized models, achieves top results in SemEval-2025 Task 4 by selectively erasing sensitive knowledge.

Embodied-Reasoner: Integrates visual search, reasoning, and action for interactive tasks, outperforming existing models in embodied environments.

ADS-Edit: Empowering autonomous driving with multimodal knowledge editing!

LookAhead Tuning: Safer LLMs via Partial Answer Previews by preserving initial token distributions.

Zero-1-to-A: Animatable avatars from a single image using video diffusion, robust to spatial & temporal inconsistencies!

MotionStreamer: Streaming motion generation w/ diffusion-based autoregressive model in causal latent space.

Creation-MMBench: Assessing Context-Aware Creative Intelligence in MLLMs

DreamRenderer: Taming attribute control in large-scale text-to-image models with a plug-and-play, training-free approach for enhanced content creation.

MagicID: ID-consistent & dynamic-preserved video customization via hybrid preference optimization.

ReCamMaster: Re-shoots videos via generative rendering, controlling camera movement from a single source, for novel perspectives and enhanced video creation.

DICEPTION: A generalist diffusion model for visual perceptual tasks.

LORAM: Train small, infer large LLMs by memory-efficient LoRA training. Enables 70B parameter model training on a 20G HBM GPU, replacing A100-80G. Reduces parameter storage cost by 15.81x.

LLMs’ knowledge acquisition is unveiled through the lens of evolving knowledge circuits, revealing how new knowledge integration depends on relevance to existing knowledge, exhibiting distinct phases …

DreamDPO: Revolutionizing text-to-3D generation by directly aligning outputs with human preferences via innovative preference optimization.

InfiGUIAgent, a novel multimodal GUI agent, leverages a two-stage training pipeline to achieve advanced reasoning and GUI interaction capabilities, outperforming existing models in benchmarks.

OneKE: a dockerized, schema-guided LLM agent system efficiently extracts knowledge from diverse sources, offering adaptability and robust error handling.

Orient Anything: Learning robust object orientation estimation directly from rendered 3D models, achieving state-of-the-art accuracy on real images.

Prompting unlocks 4K metric depth from low-cost LiDAR.

ZipAR accelerates autoregressive image generation by up to 91% through parallel decoding leveraging spatial locality in images, making high-resolution image generation significantly faster.

ScoreLiDAR: Distilling diffusion models for 5x faster, higher-quality 3D LiDAR scene completion!