Changqian Yu

Changqian Yu

Director of Kling Image · Kling AI

Building visual foundation models that perceive, reason, and create — in a unified architecture.

I lead Kling Image at Kling AI, architecting foundation models across visual generation, editing, and multimodal understanding. From pioneering real-time vision backbones (BiSeNet, Lite-HRNet) to scaling frontier multimodal systems (Kling Image 4.0, SkyReels), my work bridges foundational architecture design and global-scale products. PhD from HUST · CSIG Top-10 Dissertation Award · Stanford’s Top 2% Scientist.

Experience
Kling AI 2026.04 – Present
Director

As Director of Kling Image, I oversee Kling AI’s image foundation model series — including the native 8K Kling Image 4.0 (#6 globally · #1 among Chinese labs on Artificial Analysis). Current focus: unified multimodal models bridging visual understanding and generation.

Head of Algorithms

Led the Kling-Image-Omni team, shipping products that power visual generation and understanding at scale. Key launches include Kling-Image-O1 — bringing visual reasoning to image generation — and Kling-Image 3.0 & 3.0 Omni, the latest generation of Kling AI’s omni-image foundation models.

Kunlun Tech (Skywork) 2023 – 2025

Led multimodal AI research. Shipped Skywork-VL-32B, a vision-language model integrating vision encoders with large language models, and built the storyboard generation model powering cinematic shot planning in SkyReels. Also built a scalable Diffusion training pipeline (MoE) for text-to-image generation.

Meituan 2021 – 2023

Research Scientist at the Autonomous Delivery Department, developing trajectory prediction and motion planning models for the autonomous delivery fleet. The transformer-based prediction model was deployed on real vehicles serving millions of orders.

Huazhong University of Science and Technology 2016 – 2021
PhD Student

Focused on semantic and panoptic segmentation. Won 1st place in the COCO & Mapillary Panoptic Segmentation Challenge 2018 at ECCV. Built TorchSeg, a widely-used PyTorch segmentation codebase (1,400+ GitHub stars).

Signature Models From real-time perception to multimodal generation

Kling Image 4.0

Preview · Nov 2026
Unified Visual Generation & Editing · Kling AI

Next-generation native 8K foundation model unifying ultra-high-definition text-to-image synthesis and instruction-driven image editing. Debuted at #6 globally on both Artificial Analysis leaderboards — #1 among Chinese labs, jumping +255 Elo in Text-to-Image and +228 Elo in Image Editing over 3.0 Omni.

  • Native 8K Ultra-HD
  • Multi-Reference Consistency
  • Typography & Layout
  • Instruction-Driven Editing
Omni-Image Foundation · Kling AI

Production omni-image foundation models supporting native multi-reference synthesis, subject consistency, and controllable editing.

Kling-Image-O1 CoT Generation
Visual Reasoning · Kling AI

Reasoning-augmented image generation model introducing structured chain-of-thought planning into complex visual creation.

Video Foundation & Storyboard · Skywork

Open-source human-centric video foundation model, cinematic storyboard shot planner, and expressive portrait animation.

Skywork-VL-32B

32B VLM
Vision-Language Model · Skywork

Large-scale multimodal model integrating high-resolution vision encoders with LLMs for visual perception and reasoning.

BiSeNet V1 & V2 6,450+ Citations
Real-Time Segmentation · ECCV & IJCV

Canonical bilateral real-time segmentation architecture in academia and industry; ECCV Top-10 Influential & ESI Highly Cited.

Lite-HRNet 800+ Cites · 920+ ⭐
Lightweight High-Resolution · CVPR & MSRA

Efficient high-resolution representation backbone introducing conditional channel weighting for dense prediction.

Selected Publications View all →
LongBanana: An Expert-Verified Benchmark for Long-Context Multi-Reference Image Synthesis
Haoxiang Cao, Yuxuan Zhang, Penghui Du, Bo Li, Jiajiong Cao, Yingying Fan, Zimeng Wu, Yifu Luo, Mingzhe Zheng, Junwen Miao, Huaisong Zhang, Yang Yao, Yu Wang, Dongfu Jiang, Ping Nie, Wenhu Chen, Chaoqun Wang, Changqian Yu, Kelsey R. Allen
NeurIPS arXiv Project
StableVQ: Practical Guidelines for Stable Vector-Quantized Tokenizer Training
Bao Tang, Jiahao Guo, Haoxiang Cao, Wenyu Liu, Changqian Yu, Kun Gai, Xinggang Wang
NeurIPS 1 citations arXiv Project
Where, What, Why, and Importance: Structured Defect Grounding for Text-to-Image Feedback
Huaisong Zhang, Hao Yu, Yuxuan Zhang, Jiahe Wang, Xinrui Chen, Haoxiang Cao, Feng Lu, Wendong Zhang, Changqian Yu, Chun Yuan
NeurIPS 4 citations arXiv
Scaling Diffusion Transformers to 16 Billion Parameters
Zhengcong Fei, Mingyuan Fan, Changqian Yu, Debang Li, Junshi Huang
arXiv 73 citations arXiv Code
BiSeNet V2: Bilateral Network with Guided Aggregation for Real-time Semantic Segmentation
Changqian Yu, Changxin Gao†, Jingbo Wang, Gang Yu, Chunhua Shen, Nong Sang
IJCV ESI Highly Cited Paper 2317 citations Paper arXiv Code
BiSeNet: Bilateral Segmentation Network for Real-time Semantic Segmentation
Changqian Yu*, Jingbo Wang*, Chao Peng, Changxin Gao†, Gang Yu, Nong Sang
ECCV ECCV 2018 Top-10 Influential Papers 4142 citations Paper arXiv Code
Open Source
  • Human-Centric Video Foundation Model. 2.7k ⭐
  • Expressive Portrait Animation in Video Diffusion Transformers. 500+ ⭐
  • A Lightweight High-Resolution Network. 900+ ⭐
  • PyTorch semantic segmentation codebase — BiSeNet, DFN, DenseASPP and more. 1.4k ⭐
Research Vision & Open Roles Kling Image · Beijing

At Kling Image, we believe the next leap in visual intelligence comes from dissolving the boundary between perception, reasoning, and synthesis — scaling unified foundation models that don’t just render pixels, but understand and act in the visual world.

01

Unified Understanding & Generation

Co-designing autoregressive reasoning and flow/diffusion synthesis in a single architecture where deep comprehension sharpens controllable creation.

02

Visual Tokenizers & Representation

Building high-compression, semantically aligned 1D/2D visual tokenizers (sparse routing & stable quantization) to unlock scaling laws.

03

Visual Agentic Intelligence

Advancing rubric-grounded RL, self-evolving post-training, and tool-augmented visual agents capable of planning, critique, and refinement.

04

Multimodal Data & Scaling Infra

Engineering aesthetic- and knowledge-dense data engines alongside ultra-large Sparse MoE training and low-latency inference systems.