Text prompts; a still image; fill in missing frames of, or extend, an existing video (The Verge)
Access
Red teamers assessing harms, plus selected visual artists, designers and filmmakers (The Verge); not on sale
OpenAI's own announcement page is behind Cloudflare, so figures come from press reports of it. Ars Technica says it had not tested the model; the access and input rows are from The Verge (https://www.theverge.com/2024/2/15/24074151/openai-sora-text-to-video-ai). No price or render time was published. · 来源
Video model rubric2026 criteria
/5
Estimate from the version's published specs and reviews, 0-5; scores are only comparable within one criteria year. Scored as previewed (Feb 2024 research preview, never sold): 1920x1080 and up to 60 s per Ars Technica, text and image input plus extending or filling in an existing video per The Verge, silent per Ars ('no synchronized sound yet'). Speed is null: OpenAI gave no render time and Wired was told only 'going out for a burrito'. Realism is from curated OpenAI demo clips only, since no outside reviewer could enter prompts; all five reviews praise the look and note artifacts or cherry-picking, so 4 is capped by that caveat.[1]
Reviews
OpenAI’s Sora Turns AI Prompts Into Photorealistic Videos · wired.com · 2024年2月15日OpenAI says that Sora is distinguished by its striking photorealism—something I haven’t seen in its competitors—and its ability to produce longer clips than the brief snippets other models typically do, up to one minute.
OpenAI collapses media reality with Sora, a photorealistic AI video generator · arstechnica.com · 2024年2月16日It’s high-resolution (1920×1080), can generate video with temporal consistency (maintaining the same subject over time) that lasts up to 60 seconds, and appears to follow text prompts with a great deal of fidelity.
Steven Levy saw four demo clips (the longest 17 seconds), calls the photorealism something he had not seen in competitors, notes a Tokyo scene whose sidewalk dead-ends, and says researchers would only describe render time as nearer 'going out for a burrito' than days.[1]
Says Sora is 1920x1080, holds a subject consistent for up to 60 seconds, extends videos or animates stills, and has no synchronized sound yet; frames it as a diffusion transformer built on DALL-E 3 style recaptioning.[1]
Reports clips up to a minute, Sam Altman generating viewer-prompted videos on X, an OpenAI spokesperson saying the model will not be broadly available soon, and the weaknesses OpenAI listed (physics, cause and effect, a person running the wrong way on a treadmill).[1]
Says OpenAI briefed it under embargo, has no plan to release Sora, started third-party safety testing, and that the samples were probably cherry-picked; Tim Brooks admits people who leave the frame for long 'won't come back'.[1]
