Yizeng Han

DyDiT++: Diffusion Transformers with Timestep and Spatial Dynamics for Efficient Visual Generation

We extend DyDiT to T2I (DyFLUX) and video generation. Moreover, LoRA finetuning is supported.

RAPID^3: Tri-Level Reinforced Acceleration Policies for Diffusion Transformer

A Stitch in Time Saves Nine: Small VLM is a Precise Guidance for accelerating Large VLMs

BIFE: Better Interaction, Fewer Errors for Minute-Long Video Generation

BIFE combines sparse KV cache retrieval and block forcing to preserve long-range interactions and reduce drift in minute-long video generation.

ClinFusion: A Vision-Centric Multimodal LLM System for Holistic Medical Understanding

ClinFusion combines native 2D and 3D medical image understanding in a unified vision encoder and introduces region-grounded evaluation for clinical reports.

Agent-as-a-Router: Agentic Model Routing for Coding Tasks

We formulate model routing as an agentic Context-Action-Feedback loop and introduce ACRouter and CodeRouterBench for efficient routing across frontier LLMs on coding tasks.

K-Forcing: Joint Next-K-Token Decoding via Push-Forward Language Modeling

K-Forcing learns a push-forward language model for joint next-k-token decoding, achieving about 2.4-3.5x batch-serving speedup while reusing the autoregressive backbone.

Towards 3D-Aware Video Diffusion Models: Render-Free Human Motion Control with Mesh Tokenization

We condition video diffusion models directly on compressed 3D human mesh tokens for render-free motion control with improved 3D structure and viewpoint awareness.

Aligning Vision-Language Model With Fine-Grained Semantics for Open-Vocabulary Segmentation

FOV-Seg aligns pixel-level visual features with word-level semantics through masked text reconstruction and text-guided visual encoding.