News

Dec 2024
We release CausVid, a fast autoregressive video diffusion model that achieves state-of-the-art video quality and speed. To our knowledge, this is the first causal video generation method that competes with bidirectional diffusion in terms of quality. I hope it will open up new possibilities for applications in robotic learning, game rendering, streaming video editing, and other scenarios that require real-time and long-horizon video generation. See this twitter thread for an overview.
Oct 2024
Adobe launched a fast mode for their generative AI, reducing latency and cost by an order of magnitude, starting with the text-to-image model in Firefly. See Fast Mode for details. I led the research behind this innovation.
Sep 2024
DMD2 is accepted at NeurIPS as an oral presentation. We have one more follow-up work on the way. Stay tuned!
May 2024
We release DMD2, a few step image generator that achieves state-of-the-art image quality and speed. We also release the code
and pretrained model. Check it out!
Nov 2023
In our recent work Distribution Matching Distillation, we distill Stable Diffusion v1.5 into a one-step generator. The distilled model achieves comparable image quality while being 30 times faster. Feel free to contact me if you have any questions!
Feb 2022
My statement of purpose for Computer Science Ph.D. Programs report.

Technology Transfer

Firefly is Adobe's Generative AI initiative that powers AI features such as generative fill, text-to-image, video generation, and more. The October 2024 release introduces a default Fast Mode option, which reduces latency and cost by an order of magnitude. This enhancement is based in part on our DMD Series (DMD and DMD2) for diffusion distillation. It is the result of collaboration between our DMD team (see the DMD author list) and production partners, including Qiang Zhang, Jinrong Xie, Connelly Barnes, and many other contributors.

Papers

From Slow Bidirectional to Fast Autoregressive Video Diffusion Models
ArXiv 2024
Tianwei Yin*, Qiang Zhang*, Richard Zhang, Bill Freeman,
Frédo Durand, Eli Shechtman, Xun Huang (* equal contribution)
A fast video generator capable of streaming inference on a single GPU
at 9 FPS, while maintaining excellent visual quality
Improved Diffusion with Distribution Matching Distillation for Fast Image Synthesis
NeurIPS 2024 (Oral)
Tianwei Yin, Michaël Gharbi, Taesung Park, Richard Zhang, Eli Shechtman,
Frédo Durand, William T. Freeman,
Selected for oral presentation at the conference. Selection rate: 0.4% (72/15671)
One-step Diffusion with Distribution Matching Distillation
CVPR 2024
Tianwei Yin, Michaël Gharbi, Richard Zhang, Eli Shechtman, Frédo Durand,
William T. Freeman, Taesung Park
Turbo3D: Ultra-fast Text-to-3D Generation
ArXiv 2024
Hanzhe Hu, Tianwei Yin, Fujun Luan, Yiwei Hu, Hao Tan, Zexiang Xu, Sai Bi,
Shubham Tulsiani*, Kai Zhang* (* equal advising)
FastComposer: Tuning-Free Multi-Subject Image Generation with Localized Attention
IJCV
Guangxuan Xiao*, Tianwei Yin*, William T. Freeman, Frédo Durand, Song Han (* equal contribution)
Diffusion with Forward Models: Solving Stochastic Inverse Problems Without Direct Supervision
NeurIPS 2023 (Spotlight)
Ayush Tewari*, Tianwei Yin*, George Cazenavette, Semon Rezchikov, Joshua B. Tenenbaum,
Frédo Durand, William T. Freeman, Vincent Sitzmann (* equal contribution)
Global Tracking Transformers
CVPR 2022
Xingyi Zhou, Tianwei Yin, Vladlen Koltun, Philipp Krähenbühl
Center-based 3D Object Detection and Tracking
CVPR 2021
Tianwei Yin, Xingyi Zhou, Philipp Krähenbühl
One of the most influential CVPR papers in 2021, refer to here.
Multimodal Virtual Point 3D Detection
NeurIPS 2021
Tianwei Yin, Xingyi Zhou, Philipp Krähenbühl

Adapted from template.