I joined ByteDance as a research scientist in 2025 and currently focus on developing the Seedance Mini and Seedance Flash model families. I received my Ph.D. in Computer Science from the Institute of Automation, Chinese Academy of Sciences, where I conducted my doctoral research at the National Laboratory of Pattern Recognition under the supervision of Prof. Changsheng Xu and Prof. Weiming Dong. Prior to that, I earned my B.S. in Automation from Tsinghua University in 2020.
My research focuses on Video and Image generation, with an emphasis on training Diffusion Transformers, mixture-of-experts (MoE) architectures and vision-language models (VLMs).
Newest first · Includes arXiv preprints. Google Scholar ↗
NeurIPS 2026Spotlight
TL;DR: To keep token routing from fragmenting coherent video content, SplitMoE uses prototype-guided routing to separate semantic and general-purpose experts, converging faster and generating higher-quality video at the same active parameter budget.
NeurIPS 2026
ACM TOG 2026
TL;DR: To improve text alignment in diffusion-based image editing, HeadRouter routes instructions through semantically relevant attention heads and refines image and text tokens, producing more faithful edits without retraining the model.
ACM TOG 2026
arXiv 2026
TL;DR: To reduce the cost of image-to-video generation, FSVideo combines aggressive latent compression, memory across transformer layers, and a few-step upsampler, achieving competitive video quality at roughly ten times the speed of the open-source baselines it evaluates.
arXiv 2026
SIGGRAPH Asia 2026
TL;DR: To transfer the idea behind a visual metaphor, Beyond Pixels uses a schema grammar and cooperating agents to extract, adapt, and critique the underlying analogy, improving metaphor consistency and creativity over the evaluated baselines.
SIGGRAPH Asia 2026
CVPR 2026
TL;DR: To reduce interference between image generation and editing tasks, TAG-MoE trains expert routing with structured descriptions of task intent, improving output fidelity and quality over dense baselines.
CVPR 2026
SIGGRAPH 2025
TL;DR: To generate images across a visual theme without fine-tuning, IP-Prompter iteratively adjusts visual prompts from reference images, improving character identity, style consistency, and text alignment across scenes.
SIGGRAPH 2025
IEEE TVCG 2025
TL;DR: To bring new subjects into artworks while preserving their visual harmony, CreativeSynth uses cross-art attention to fuse multimodal guidance with a diffusion model, supporting artistic synthesis across painting categories.
IEEE TVCG 2025
IEEE TVCG 2025
TL;DR: To transfer motion from one reference video without copying its appearance, MotionCrafter separates temporal motion guidance from spatial appearance control, reproducing actions or camera motion with new characters and styles.
IEEE TVCG 2025
ACM TOG 2025
TL;DR: To keep content and style from becoming entangled in personalized generation, B4M learns them in separate low-rank adapter spaces and recombines them, improving alignment with the target subject, appearance, and prompt.
ACM TOG 2025
SIGGRAPH Asia 2024
TL;DR: To align generated music with dance movements, separate rhythm and genre encoders map a dance clip to tokens for a text-to-music model, improving synchronization and adapting to tempo changes.
SIGGRAPH Asia 2024
arXiv 2024
TL;DR: To control portrait lighting independently of scene content, LumiSculpt adds a module trained on the LumiHuman dataset that lets a video diffusion model follow a light source's intensity, position, and path consistently over time.
arXiv 2024
IEEE TVCG 2024
TL;DR: To make arbitrary style transfer methods easier to compare, this study builds a fine-grained benchmark that pairs objective metrics with structured human judgments and reveals where they agree or diverge across four model families.
IEEE TVCG 2024
IEEE SPL 2024
TL;DR: To stylize video from text without training a video model, Style-A-Video combines image diffusion with sampling and temporal consistency controls, retaining scene structure and reducing flicker at lower cost than prior methods.
IEEE SPL 2024
AAAI 2024
TL;DR: To transfer musical styles from scarce examples, time-varying textual inversion learns features at multiple levels of a mel spectrogram and combines them with bias-reduced sampling to transfer instrument styles and incorporate natural sounds into melodies.
AAAI 2024
IEEE TNNLS 2024
TL;DR: To balance text-driven stylization with content preservation, DiffStyler combines two diffusion processes and learns an initial noise representation from the source image, producing stylized images that retain more of the original structure.
IEEE TNNLS 2024
SIGGRAPH Asia 2023 (Journal Track) / ACM TOG
TL;DR: To edit visual attributes independently during diffusion personalization, ProSpect learns distinct prompt embeddings for different denoising stages, giving finer control over properties such as material, style, and layout.
SIGGRAPH Asia 2023 (Journal Track) / ACM TOG
ACM TOG 2023
TL;DR: To reduce style inconsistency across image stylization architectures, UCAST learns style features with an input-dependent contrastive objective, improving stylization across CNN, transformer, and flow-based models.
ACM TOG 2023
Leonardo 2023
TL;DR: To automate portrait map art, an asymmetric translation model encodes the portrait and map separately and harmonizes their features, synthesizing artworks that blend facial forms with cartographic details.
Leonardo 2023
CVPR 2023
TL;DR: To transfer artistic qualities that are difficult to describe in words, InST inverts a single painting into the conditioning space of a diffusion model, carrying its colors, shapes, and motifs into new compositions.
CVPR 2023
SIGGRAPH 2022
TL;DR: To address style inconsistency caused by fixed image statistics, CAST uses contrastive learning and a model of the broader style distribution to encode painting styles, improving transfer fidelity and reducing local distortions.
SIGGRAPH 2022
GitHub