AI People Wiki

Builders & open source

Bichen Wu

Computer vision researcher; first author of Meta's Fairy video-editing paper and of FBNet

Bichen Wu's GitHub profile describes Wu as a research scientist in AI, and the work cited here covers efficient computer vision and generative video. Wu is first author of Fairy (2023), a paper from GenAI at Meta that adapts image-editing diffusion models for fast, instruction-guided video-to-video editing. Earlier, while listed with UC Berkeley, Wu was first author of FBNet (2018), a differentiable neural architecture search method for designing accurate and efficient ConvNets for mobile devices; the paper notes that Wu's work on it was done while interning at Facebook.123

At a glance

Current role
None confirmed in the sources
Education
UC Berkeley
Work listed
2
Sources cited
10
Profile reviewed

Efficient neural networks

Much of Wu's early research looked at making neural networks small and fast enough for edge devices such as cars and phones. SqueezeDet (2016) is a compact fully convolutional network for real-time object detection in autonomous driving, and SqueezeSeg (2017) segments road objects in 3D LiDAR point clouds; Wu released implementations of both on GitHub, the SqueezeDet one in TensorFlow. Wu's 2019 PhD dissertation, "Efficient Deep Neural Networks", gathers this line of work.4516

Wu was also first author of "Shift" (2017), a zero-FLOP, zero-parameter alternative to spatial convolutions, and of FBNet (2018), which used differentiable neural architecture search to design efficient networks for mobile hardware. The FBNet family continued with FBNetV5 (2021), on searching architectures for multiple vision tasks in a single run.738

Vision transformers and generative video

Wu's 2020 paper "Visual Transformers" proposed representing images as semantic visual tokens processed with transformers. In late 2023, as first author of Fairy from GenAI at Meta, Wu worked on fast instruction-guided video editing, and Wu co-authored FlowVid, which uses optical flow together with spatial conditions to keep edited video frames consistent.9210

Work and contributions

  • Fairy: Fast Parallelized Instruction-Guided Video-to-Video Synthesis Published research

    First author of the 2023 paper from GenAI, Meta introducing Fairy, which uses anchor-based cross-frame attention to extend image-editing diffusion models to video editing.2

  • FBNet Published research

    First author of the 2018 FBNet paper proposing differentiable neural architecture search for hardware-aware, efficient ConvNets on mobile devices.3

Timeline

  1. Posted SqueezeDet, a small fully convolutional network for real-time object detection, as first author.4
  2. Posted SqueezeSeg, for real-time road-object segmentation from 3D LiDAR point clouds, as first author.5
  3. Posted "Shift", a zero-FLOP, zero-parameter alternative to spatial convolutions, as first author.7
  4. Posted FBNet, a differentiable architecture search method for efficient ConvNets, as first author.3
  5. Posted the PhD dissertation "Efficient Deep Neural Networks".6
  6. Posted "Visual Transformers", on token-based image representation, as first author.9
  7. Posted FBNetV5, on neural architecture search for multiple tasks in one run, as first author.8
  8. Posted Fairy, a fast instruction-guided video-to-video editing method from GenAI at Meta, as first author.2
  9. Co-authored FlowVid, a video-to-video synthesis framework using optical flow for temporal consistency.10

Documented roles

The sources do not confirm a current role; these are the roles they document.

  • Listed with GenAI, Meta as first author of the Fairy paper (2023); FBNet work done while interning at Facebook, Meta23

Companies in their history

  • Meta 106 people in the wiki

Suggest a correction or removal

Profiles cover public professional work only, and each statement cites a source. If something is out of date, missing or wrong, tell us and link a source. People can also ask for their profile to be corrected or removed.

References

10 sources · profile reviewed Oct 21, 2026 · data updated Oct 21, 2026

  1. BichenWuUCB (Bichen Wu) on GitHubGitHub · Official page · accessed Oct 21, 2026
  2. Fairy: Fast Parallelized Instruction-Guided Video-to-Video Synthesis (arXiv 2312.13834)arXiv · Paper · published Dec 20, 2023 · accessed Oct 21, 2026
  3. FBNet: Hardware-Aware Efficient ConvNet Design via Differentiable Neural Architecture Search (arXiv 1812.03443)arXiv · Paper · published Dec 9, 2018 · accessed Oct 21, 2026
  4. SqueezeDet: Unified, Small, Low Power Fully Convolutional Neural Networks for Real-Time Object Detection for Autonomous Driving (arXiv 1612.01051)arXiv · Paper · published Dec 4, 2016 · accessed Oct 21, 2026
  5. SqueezeSeg: Convolutional Neural Nets with Recurrent CRF for Real-Time Road-Object Segmentation from 3D LiDAR Point Cloud (arXiv 1710.07368)arXiv · Paper · published Oct 19, 2017 · accessed Oct 21, 2026
  6. Efficient Deep Neural Networks — Ph.D. dissertation of Bichen Wu (arXiv 1908.08926)arXiv · Paper · published Aug 20, 2019 · accessed Oct 21, 2026
  7. Shift: A Zero FLOP, Zero Parameter Alternative to Spatial Convolutions (arXiv 1711.08141)arXiv · Paper · published Nov 22, 2017 · accessed Oct 21, 2026
  8. FBNetV5: Neural Architecture Search for Multiple Tasks in One Run (arXiv 2111.10007)arXiv · Paper · published Nov 19, 2021 · accessed Oct 21, 2026
  9. Visual Transformers: Token-based Image Representation and Processing for Computer Vision (arXiv 2006.03677)arXiv · Paper · published Jun 5, 2020 · accessed Oct 21, 2026
  10. FlowVid: Taming Imperfect Optical Flows for Consistent Video-to-Video Synthesis (arXiv 2312.17681)arXiv · Paper · published Dec 29, 2023 · accessed Oct 21, 2026