Currently, I am a research staff at TeleAI since June 2025. Previously, I was a research expert at Bytedance since Jan 2024 and a senior engineer at Alibaba. My research focuses on multi-reference-guided image/video generation since 2022, and video understanding during 2018 to 2022. I have practical industrial experiences dealing with catastrophic forgetting problems when introducing new controllable generation capabilities to the foundation models. I created and open-sourced TeleStyle V1 and V2 series for content-preserving style transfer given content reference and style reference, achieving comparable performance with closed-source competitors and were well-recieved by open-source community (more than 1k upvotes on Reddit) and industry (integrated by fal.ai). I was the owner of style reference customization functionality in 即梦2.0 era and launched the first style code function in China which was popular among users. I got my bachelor degree from Tsinghua University in 2018.

My current main skills are:

  1. Training experiences of multi-ref image customization on vlm+connector+dit and mixtures of transformers structures.
  2. Construction and synthesizing multi-ref training datasets.
  3. Joint training of multi tasks on image customization.
  4. Training with noisy data, curriculum learning, dealing with catastrophic forgetting.
  5. Training video generation model (minimax h3) to conduct mulit-ref image customization (content-preserving style transfer).
  6. Training SOTA reference to video (R2V) models.

News

  • 2026.07 TeleStyle V2 supports SenseNova-U1, reinforces SenseNova U1 for Content-Preserving Style Transfer and preserves its general image editing capability.
  • 2026.07 Act2Cut is accepted by SIGGRAPH Aisa 2026.
  • 2026.06 TeleStyleV2 has been integrated by fal.ai at [fal-ai/telestyle-v2] .
  • 2026.06 TeleStyleV2 has been integrated by mlx-community at [mlx-community/TeleStyleV2-Qwen-Image-Edit-2511-bf16] .
  • 2026.03 I release the technical report, project, code and demo of TeleStyleV2.
  • 2026.03 I finished training TeleStyleV2.
  • 2026.01 TeleStyleV1 got more than 1k upvotes on reddit community.

Selected Papers

[1] TeleStyle V2: Beyond Content-Preserving Style Transfer with Self-Distillation and Distribution-Matching-Distillation
Shiwen Zhang, Yifan Xu, et al
Tech Report
Paper Code GitHub stars Project Model Demo BibTeX
[2] TeleStyle: Content-Preserving Style Transfer in Images and Videos
Shiwen Zhang, Xiaoyan Yang, Bojia Zi, et al
Tech Report
Paper Code GitHub stars Model Demo BibTeX
[3] QwenStyle: Content-Preserving Style Transfer with Qwen-Image-Edit
Shiwen Zhang, et al
Tech Report
Paper Code GitHub stars Model Demo BibTeX
[4] DreamO: A Unified Framework for Image Customization
Chong Mou, Yanze Wu, ..., Shiwen Zhang,..., et al
SIGGRAPH Asia 2025
Paper Code GitHub stars Model Demo
[5] Act2Cut: Continuous Multi-Shot Video Narrative Match on Action-Cut
Cailin Zhuang, Yaoqi Hu, Zheng Dong, Shiwen Zhang et al
SIGGRAPH Asia 2026
Paper
[6] CDST: Color Disentangled Style Transfer for universal style transfer tasks
Shiwen Zhang, Zhuowei Chen, Lang Chen, Yanze Wu
Tech Report
Paper
[7]Fast Imagic: Solving Overfitting in Text-guided Image Editing via Disentangled UNet with Forgetting Mechanism and Unified Vision-Language Optimization
Shiwen Zhang
NeurIPS 2024 WorkShop
PMLR 2024
Paper Code GitHub stars
[8] Forgedit: Text guided Image Editing with Diffusion Models via Learning and Forgetting
Shiwen Zhang, Shuai Xiao, Weilin Huang
Tech Report
Paper Code GitHub stars
[9] TFCNet: Temporal Fully Connected Networks for static unbiased temporal reasoning
Shiwen Zhang
Tech Report
Paper
[10] V4D: 4D Convolutional Neural Networks for Video-level Representation Learning
Shiwen Zhang, Sheng Guo, Weilin Huang, Matthew R. Scott and Limin Wang.
ICLR 2020
Paper Code GitHub stars
[11] Knowledge Integration Networks for Action Recognition
Shiwen Zhang, Sheng Guo, Limin Wang, Weilin Huang and Matthew R. Scott.
AAAI 2020
Paper

Services

Reviewer of ICLR, NeurIPS, ICML, CVPR, ICCV, ECCV, SIGGRAPH Asia.