Brooks, Peebles, Holmes, DePue, Guo, Jing, Schnurr, Taylor, Luhman,
Luhman, Ng, Wang and Ramesh, OpenAI (February 2024). This is a web page
with no arXiv id or PDF, and url: is the canonical page. openai.com
refuses this build environment (403), and web.archive.org and archive.ph
refuse the connection. It was read in full through the r.jina.ai reader
rendering of the canonical URL, retrieved 2026-09-24. That copy includes
the author list and all 32 references (reference 30, DALL-E 3, renders
empty). The embedded videos are not in it, so the claims below that rest
on videos are reported as claims.
Key takeaways
- The architecture, as far as it is given. A network compresses video “both temporally and spatially” into a latent, with a decoder back to pixels. The latent is cut into spacetime patches that act as transformer tokens, and an image is a one-frame video. Sora “is a diffusion transformer”, citing DiT (LIT-448). It is trained to predict the “clean” patches from noisy ones. The report says “Model and implementation details are not included”. No compression factor, patch size, parameter count, noise schedule, objective parametrization, data size or attention layout is given.
- Native size instead of crops. The report trains at native duration, resolution and aspect ratio. The reason given is sampling flexibility (1920x1080 to 1080x1920, and quick prototyping at low size) and better framing. At inference the output size is set by the grid of noise patches, and images up to 2048x2048 are generated the same way.
- Video recaptioning, and prompt upsampling at inference. “We apply the re-captioning technique introduced in DALL·E 3 to videos”. A “highly descriptive captioner model” captions every training video, and GPT expands short user prompts into long captions. The report does not say whether the captioner sees video or frames.
- Conditioning on images and videos. It animates DALL-E images, extends videos backward and forward (and into loops), applies SDEdit for zero-shot video-to-video editing, and interpolates between two videos.
- “Emerging simulation capabilities”. 3D consistency under camera motion, object permanence through occlusion, simple state changes (brush strokes, bite marks), and Minecraft rendered with “a basic policy”. These “emerge without any explicit inductive biases … purely phenomena of scale”.
Where the hedges are
Per DP-010:
- There is no measurement anywhere in it. The report has no table, metric, number of samples or baseline. “Sample quality improves markedly as training compute increases” is supported by one set of samples at three compute levels (base, 4x, 32x), with seeds and inputs fixed. The claim that native aspect ratios improve framing compares Sora with a square-crop variant on examples. Neither comparison is quantified, and neither states that anything else was held fixed.
- “Purely phenomena of scale” is asserted, not tested. No smaller model is shown failing at the capabilities listed, and no model with a 3D or object prior is compared. The Discussion lists what fails (glass shattering, eating, incoherence over long samples, objects appearing spontaneously) and points to the landing page for more. Both the successes and the failures are sampled examples, not rates.
- “Improves text fidelity as well as the overall quality” is the recaptioning result, and it is one sentence with no evidence shown. DALL-E 3 (LIT-648), which it cites for the technique, measured CLIP score in a controlled setting. This report measures nothing.
- Later papers attribute details to it that it does not contain. LTX-Video (LIT-618, §2) cites this report as proposing “an upsampler working directly in pixel space, conditioned on the latent outputs of the base model”. This report says nothing about an upsampler.
Standing in the anthology
The video reports filed after it cite it, and several benchmark against Sora’s outputs. Movie Gen (LIT-626) compares with Sora in a human study, where the Sora arm used a different recipe. Wan (LIT-619) scores it on VBench and on Wan-Bench, and Open-Sora 2.0 (LIT-634) reports its VBench gap to it. Its standing comes from being the reference point others measure against. Per DP-005, that is adoption, and it is not evidence for any practice.
What it adopts, the record already holds as practices with better sources:
- Joint image and video training is SOTA-386. Sora adopts it and does not ablate it.
- Descriptive recaptioning is SOTA-389. Sora adopts it, but its captioner is not described as watching video, so it is not an adopter of that practice’s specific claim.
- Native-size training over fixed crops has no practice for generative models. Sora’s square-crop comparison is qualitative. NaViT (LIT-657), its reference 18, is now filed. It supports SOTA-398, a measured practice for ViT pretraining, which isolates aspect ratio only in a probe.
Its lasting contribution is framing. It presents spacetime-patch diffusion transformers as the scalable video representation and scaling as the route to “world simulators”. The report asserts both and demonstrates neither.
Filed without a NOTE. This note is itself the full reading: there are
about 1,800 words of text before the references, and the evidence is in videos this reading
could not see.