Xiao Yang

Research Director at Canva

Email/ X/ LinkedIn/ Google Scholar

About

I am a Research Director at Canva on Efficient AI (Image/Video/LLMs). My goal is making Generative AI more efficient and accessible, for both inference and training.

Prior to that I worked at Bytedance as a Senior Staff Manager, managing a team of 10+ people on AIGC foundation model development and applications. Before that I also worked at Meta shortly on video advertisement products.

I got my Ph.D degree from Computer Science department, UNC Chapel Hill, advised by Dr. Marc Niethammer; I obtained my bachelor degree in Electrical Engineering from Tsinghua University.

My past works spans across various aspect of generative models, including:

Efficient model: FlashGaze, Helios, FSVideo, SDXL-Lightning, AnimateDiff-Lightning
Foundation models: SplitMoE, Seeddream, zero-terminal SNR, MVDream
Applications: MagicPose, IP-Prompter, and 20+ high-profile product launches (such as this one and this one).

News

Hiring! I am hiring people interested in efficient AI (distillation, quantlization, efficient architecture design, training efficiency improvement etc) for image/video/LLM models in China. Please feel free to contact me if you are interested!
Introducing Canva Magic Layers: transforming flat images into editable, layered designs. My team is responsible for end-to-end acceleration of the underlying models. [Canva Newsroom]

Selected Publications

Breaking the Uniformity Trap: Scaling Video Diffusion Model via SplitMoE.

NeurIPS 2026 · Spotlight

A 27B video diffusion MOE model with 14B activated parameters, splitting semantic and generic experts to improve generation quality and routing coherence.

Project Paper
FlashGaze: Training-Free Multi-Scale Patch Pruning For Efficient Video Understanding.

arXiv 2026

Training-free patch pruning for efficient video understanding. On Qwen3-VL-8B, retains 98% of baseline accuracy on LongVideoBench with up to 5.4× faster vision encoding and 17× faster MLLM prefill.

Paper
Helios: Real Real-Time Long Video Generation Model.

arXiv 2026

A 14B real-time long video generation model capable of generating video at 19.5 FPS on a single H100.

Project Paper Code
FSVideo: Fast Speed Video Diffusion Model in a Highly-Compressed Latent Space.

arXiv 2026

A 28B video generation model achieving high quality, 720P video generation while being 42.3× faster than WAN 14B video models.

Project Paper
IP-Prompter: Training-Free Theme-Specific Image Generation via Dynamic Visual Prompting.

SIGGRAPH 2025

Training-free personalization method achiving state-of-the-art character identity preserving, style consistency and text alignment results.

Project Paper Code
SDXL-Lightning: Progressive Adversarial Diffusion Distillation.

arXiv 2024

Industry-standard few-step SDXL acceleration method.

Paper Model
AnimateDiff-Lightning: Cross-Model Diffusion Distillation.

arXiv 2024

SOTA few-step video generation for animatediff; >60 million downloads on Huggingface.

Paper Model
MVDream: Multi-view Diffusion for 3D Generation.

ICLR 2024

First method to solve Janus problem for 3D generation.

Project Paper Code
MagicPose: Realistic Human Poses and Facial Expressions Retargeting with Identity-aware Diffusion.

ICML 2024

High performance pose and expression retargeting.

Project Paper Code
Common Diffusion Noise Schedules and Sample Steps are Flawed.

WACV 2024

Zero-terminal SNR. Widely-used noise schedule fix for diffusion model for image/video/3D generation

Paper
MoMA: Multimodal LLM Adapter for Fast Personalized Image Generation.

ECCV 2024

Training-free image personalization method via multimodal LLM prompting.

Project Paper Code
PAniC-3D: Stylized Single-view 3D Reconstruction from Portraits of Anime Characters.

CVPR 2023

GAN-based single-view-to-3D reconstruction method for anime characters with strong stylization handling capability.

Paper Code
Shifted Diffusion for Text-to-image Generation.

CVPR 2023

SOTA open-source text-to-image generation model in DALL-E 2 era.

Paper Code
SemanticStyleGAN: Learning Compositonal Generative Priors for Controllable Image Synthesis and Editing.

CVPR 2022

Stylegan-based model for compositional image generation and editing.

Project Paper Code

Last update date: 2026/10.