§ Publications — Research Snapshot
My research centers on Multi-modal Agentic RL, Unified Models, and Reward Modeling — designing large-scale post-training recipes that align unified generation models with human preferences, and building agents that reason, retrieve, and self-evolve by interacting with real-world tools and visual experience. My work has long lived in visual generation: GenEvolve pushes agentic visual reasoning over unified models; PosterCraft, PosterOmni, and PosterReward turn unified generation and reward-driven post-training into dependable creative tools.
arXiv 2026
A self-evolving image-generation agent that orchestrates external search, visual references, and generation knowledge into prompt-reference programs, then improves through tool-orchestrated visual experience distillation.
CVPR 2026
One model for poster creation—unifying local edits and global design for generalized multi-task image/poster-to-poster generation. Your intelligent assistant for high-quality aesthetic poster creation!
CVPR 2026
ICLR 2026
CVPR 2025
A self-evolving image-generation agent that orchestrates search, visual references, and internal generation knowledge into prompt-reference programs, learning better tool use through visual experience distillation.
One model for poster creation—unifying local edits and global design for generalized multi-task image/poster-to-poster generation. Your intelligent assistant for high-quality aesthetic poster creation!
A comprehensive reward model for design aesthetics and typography, trained on automated preference dataset to unlock accurate evaluation for high-quality graphic design generation.
Proposes a one-step, reference-controllable haze generator that better matches real-world haze complexity than classic physics pipelines, creating high-quality paired data to boost real-world dehazing performance.
Provides a systematic benchmark study of GPT-4o image generation across 20+ tasks (text-to-image, image-to-image, image-to-3D, image-to-X), summarizing strengths/limitations and what they imply for unified multimodal generation.
A multimodal agent reasons about adverse scenes, calling specialized restoration experts to stabilize perception stacks for autonomous vehicles.
Uses an MMPO/DPO-enhanced MLLM as an evaluator to rank/filter pseudo-labels, enabling semi-supervised training that improves real-world desnowing without relying on dense paired GT.
Builds an iterative data engine to harvest large-scale unlabeled data and progressively select high-reliability pseudo labels, improving mirror detection robustness and generalization in diverse scenes.
Introduces a depth-prompting paradigm: leverages Depth Anything’s stable depth priors as prompts to guide dehazing, aiming for plug-and-play real-world dehazing under complex haze.
Proposes a training-free, unsupervised diffusion guidance framework that steers a pretrained diffusion model using attribute-based guidance for effective real-world low-light enhancement.
Proposes an MLLM-driven agent that diagnoses degradations and plans a task sequence + selects expert models from a tool/model pool to restore images.
Using a prompt pool to compose weather-specific prompts on the fly, plus Depth-Anything–constrained scene prompts to stabilize background reconstruction under unseen adverse-weather combinations.
Presents DTPM (Diffusion Texture Prior Model) that explicitly models high-quality texture priors to preserve fine details and structure when applying diffusion to restoration.
A transformer that uses sparse sampling attention to model global rain degradation relations and an uncertainty-driven ranking strategy to focus on hard-to-restore regions for unified deraining.
Proposes an efficient single image desnowing framework by modeling uncertainty in degradation perception and background reconstruction.