Adapting pretrained generative models to robot control through current-observation denoising, without a separate future-target branch.

Introducing ReChannel, a lightweight framework that repurposes pretrained DiT for dense prediction by mapping each patch token directly to task-native pixel fields instead of casting dense outputs as image-like RGB targets through a generative VAE decoder.

investigate the underlying mechanism behind the evolution from SRA to Self-Flow, showing that the methodological improvement, as well as the resulting performance gain, mainly stems from data augmentation.

Introducing Z-Reward, a teacher-student frame work for visual reward modeling that represents human preference as a reasoning-conditioned score distribution rather than a single deterministic scalar.

Introducing Z-Image-Turbo ++, a 2-step image generation model trained with Distribution-Aligned Adversarial Learning, Step-Decoupled Parameterization, and End-to-End Training with Iterative Regularization.

Showing that explicit timestep embeddings of diffusion models can be replaced by geometry, and implement two practical mechanisms: schedule-level manifold decoupling and orthogonal time-space disentanglement.

Proposing a precise manifold-aware SDE for exploration, dual trust-region KL regularization and gradient norm equalization for better video diffusion model's post-training.

Demonstrating that video generation models are not merely for visual synthesis but also as powerful engines for visual reasoning.

A lightweight framework that leverages off-the-shelf VAE's features to addresses extra forward overhead in the representation alignment for diffusion transformers.

A novel training framework that facilitates MLLMs in learning more discriminative visual representations via masked image modeling in the joint latent semantic space of LLM.

An End-to-End Virtual Try-on model that directly fits the target garment onto the person image.

Reformulating RVOS as a continuous, text-conditioned flow from video to mask and achieving leading performance.

Transferring SAM to affordance grounding task and showing robust performance for both seen and unseen actions.

Adapting Grounding DINO for STVG using a novel PEFT strategy with specialized adapters and a dedicated temporal decoder.

A novel framework for semantic segmentation dataset generation promotion by jointly diffusing images with pixel-level annotations.

The LLM-based agent which achieving leading machine learning engineering capability by combining a curated ML knowledge base with proposed Monte Carlo Graph Search.

A low-biased general annotated dataset (e.g, ImageNet) generation framework helps to obtain more generalized visual backbones.