Bill Peebles
Last reviewed
Sources
10 citations
Review status
Source-backed
Revision
v1 · 1,685 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
10 citations
Review status
Source-backed
Revision
v1 · 1,685 words
Add missing citations, update stale details, or suggest a clearer explanation.
William "Bill" Peebles is a computer scientist and machine learning researcher best known for co-inventing the Diffusion Transformer (DiT) architecture and for leading the development of Sora, OpenAI's text-to-video generation system. With Saining Xie, Peebles introduced DiT in 2022, replacing the convolutional U-Net backbone of earlier diffusion models with a pure transformer. The design became a foundation for a generation of image and video generators, Sora among them. Peebles joined OpenAI in 2023, co-led and then headed the Sora team, and oversaw the September 2025 launch of Sora 2 and its companion social app before leaving the company in April 2026 as OpenAI wound the product down.[1][8]
Peebles studied at the Massachusetts Institute of Technology (MIT) from 2015 to 2019, where he worked on computer vision and generative modeling in the lab of Antonio Torralba through MIT's SuperUROP undergraduate research program. His earliest publications date from this period, including "Seeing What a GAN Cannot Generate" (ICCV 2019) and "Semantic Photo Manipulation with a Generative Image Prior" (SIGGRAPH 2019), both rooted in the analysis and editing of images produced by generative adversarial networks.[1]
He then moved to the University of California, Berkeley for doctoral study, joining the Berkeley Artificial Intelligence Research (BAIR) lab from 2019 to 2023 under the supervision of Alexei "Alyosha" Efros. During his PhD he was supported by a National Science Foundation Graduate Research Fellowship and completed internships at Meta's Fundamental AI Research lab (FAIR), Adobe Research, and NVIDIA. His dissertation, "Generative Models of Images and Neural Networks," was filed in May 2023 and tied together three strands of his work: diffusion transformers, generative models trained on neural network checkpoints, and the use of pretrained GANs as a source of supervision for visual correspondence.[2]
Peebles's graduate research centered on scaling generative models and extending them beyond their original domains. Before DiT, his most recognized result was "GAN-Supervised Dense Visual Alignment" (CVPR 2022), which used a pretrained GAN generator to learn dense correspondences across images without human annotation; the paper was named a Best Paper Finalist. He also explored meta-learning with "Learning to Learn with Generative Models of Neural Network Checkpoints," a system that trained a diffusion model over hundreds of thousands of optimizer trajectories so that it could sample neural network parameters meeting a target loss.[1][2]
The work that made Peebles widely known is "Scalable Diffusion Models with Transformers," released on arXiv in December 2022 and co-authored with Saining Xie. The paper introduced the Diffusion Transformer, or DiT, a class of latent diffusion models that discards the U-Net architecture standard in image generators and instead processes sequences of latent image patches with a transformer. Conditioning information such as the diffusion timestep and class label is injected through an adaptive layer normalization scheme the authors call adaLN-Zero.[3]
DiT's central finding was a clean scaling relationship: as the transformer's compute, measured in Gflops, increases, sample quality improves predictably. The largest model, DiT-XL/2, set a new state of the art on class-conditional ImageNet generation, reaching a Frechet Inception Distance (FID) of 2.27 at 256x256 resolution. The result demonstrated that the transformer scaling behavior already familiar from large language models also held for diffusion-based image synthesis.[3]
The paper had a notably rocky start. It was rejected from CVPR 2023 on the grounds of "limited novelty," with reviewers describing it as a straightforward combination of a transformer and a latent diffusion model. It was subsequently accepted to the IEEE/CVF International Conference on Computer Vision (ICCV) 2023 as an oral presentation. Xie later cited the episode publicly as evidence that peer review struggles to predict impact: DiT went on to underpin OpenAI's Sora and, through the related SiT formulation, influenced systems such as Stable Diffusion 3.[3][4]
The table below summarizes Peebles's principal publications.
| Year | Work | Venue | Note |
|---|---|---|---|
| 2019 | Seeing What a GAN Cannot Generate | ICCV | Oral |
| 2019 | Semantic Photo Manipulation with a Generative Image Prior | SIGGRAPH | |
| 2020 | The Hessian Penalty | ECCV | Spotlight |
| 2022 | GAN-Supervised Dense Visual Alignment | CVPR | Best Paper Finalist |
| 2022 | Learning to Learn with Generative Models of Neural Network Checkpoints | arXiv | |
| 2022 | Scalable Diffusion Models with Transformers (DiT) | ICCV 2023 | Oral |
Peebles joined OpenAI in 2023 to work on video generation, bringing the DiT recipe into the video domain. He co-led the Sora effort with Tim Brooks, a former Berkeley colleague who had helped start the project in early 2023. The two engineered a model that represented video as sequences of spacetime latent patches, allowing a single diffusion transformer to generate clips of varying resolution, duration, and aspect ratio.[5]
OpenAI first previewed Sora on February 15, 2024, accompanied by a technical report, "Video generation models as world simulators," that framed large-scale video models as a path toward general-purpose simulators of the physical world. Brooks left OpenAI for Google DeepMind in October 2024 to work on video and world simulation, after which Peebles continued as head of the Sora team. Sora became publicly available to ChatGPT Plus and Pro subscribers in the United States and Canada on December 9, 2024.[5][6][9]
Under Peebles, OpenAI released Sora 2 on September 30, 2025, alongside a standalone mobile application built around a vertical, TikTok-style feed of AI-generated clips. The app's signature feature, Cameo, let users insert a short verified selfie video of themselves, or of friends who granted permission, into generated scenes. The launch was an immediate consumer hit: Peebles announced that the Sora app surpassed one million downloads in fewer than five days, faster than ChatGPT had at its own debut, and it climbed to the top of Apple's App Store, where it remained for weeks.[7][8]
The app's success was short-lived as a business. Reports placed Sora's operating cost at roughly one million dollars per day in compute, and the product drew intellectual-property complaints over generated likenesses and copyrighted material. On March 24, 2026, OpenAI announced it would discontinue Sora as a standalone product. The web and app experiences shut down on April 26, 2026, and the company scheduled the Sora API for discontinuation on September 24, 2026. The underlying Sora 2 model was retained inside ChatGPT, and OpenAI redirected much of the Sora research team toward world-simulation work aimed at robotics.[8][10]
Peebles announced his departure from OpenAI on Friday, April 17, 2026, on the same day as two other senior figures: Kevin Weil, the vice president who had led OpenAI for Science, and Srinivas Narayanan, an executive in the company's enterprise applications group. Commentators framed the simultaneous exits as part of a broader strategic shift in which OpenAI pared back consumer-facing "side quests" to focus on enterprise products and core research ahead of a potential public offering.[8][9]
In the note he shared with his team and posted publicly, Peebles called building Sora "zero-to-one" with his colleagues "the honor and adventure of a lifetime." He argued that Sora had spurred a large wave of investment in video across the industry and that ambitious research of that kind needs room to operate apart from a company's main lines of business, writing that "cultivating entropy is the only way for a research lab to thrive long-term." OpenAI did not announce a successor for the Sora franchise, and Peebles did not immediately disclose his next venture.[8][9]
The timeline below captures the key dates of Peebles's public career.
| Date | Event |
|---|---|
| 2015-2019 | Undergraduate study at MIT (Torralba lab) |
| 2019-2023 | PhD at UC Berkeley (BAIR, advised by Efros) |
| Dec 2022 | DiT paper released with Saining Xie |
| 2023 | Joins OpenAI; co-leads Sora with Tim Brooks |
| Feb 15, 2024 | Sora previewed |
| Dec 9, 2024 | Sora launched to ChatGPT Plus and Pro users |
| Sep 30, 2025 | Sora 2 and Sora app released |
| Mar 24, 2026 | OpenAI announces Sora discontinuation |
| Apr 17, 2026 | Peebles announces departure from OpenAI |
| Apr 26, 2026 | Sora web and app shut down |
Peebles is regarded as one of the architects of the transformer-based diffusion models that came to dominate generative media in the mid-2020s. The Diffusion Transformer is widely cited and has been adopted or adapted across a range of open and commercial image and video systems, and its initial rejection followed by outsized influence is frequently used as a cautionary example about the limits of academic peer review. His earlier "GAN-Supervised Dense Visual Alignment" was a CVPR 2022 Best Paper Finalist, and his doctoral research was supported by an NSF Graduate Research Fellowship. Beyond the research itself, his stewardship of Sora placed him at the center of the commercial emergence of consumer AI video between 2024 and 2026.[1][3][8]