How to tell if an image is Sora (OpenAI)
Sora is OpenAI's text-to-video model. It was video-first, so most stills people check are frame grabs, screenshots, or reuploads rather than files a generator handed someone directly. OpenAI announced Sora's discontinuation on 24 March 2026, closed the Sora web and mobile apps on 26 April 2026, and scheduled the developer API to shut down on 24 September 2026, so Sora material now circulating is archived or re-shared rather than newly made in the app. That matters for identification: a frame pulled out of a video carries almost none of the provenance the original file had, and it carries video compression that a still from an image generator does not.
Visual tells
A visible moving watermark on standard app downloads: an animated white cloud logo that changes position and opacity between frames, and in many exports an account username alongside it. Because it moves, a sequence of grabs from one clip shows it in different places. Its absence means nothing, since Pro-tier downloads came without it and cropping or removal tools were widely available within about a week of Sora 2's release.
Video compression signatures that a native still generator would not produce: blocky 8x8 or 16x16 artefacts in flat areas, colour smearing around high-contrast edges, and directional motion blur on moving subjects. A frame grab is a re-encoded video frame, and it looks like one under magnification.
Frozen mid-transition failures. Sora's known weaknesses are temporal, so a frame pulled from the wrong moment catches them in progress: smeared or duplicated fingers and limbs, objects partly formed or partly dissolved, and edges that trail rather than resolve. Frames chosen as thumbnails or hero shots usually avoid this, so scan the unglamorous frames.
Contact and physics errors of the kind OpenAI itself acknowledged, describing limited capacity to simulate complex physics, to understand causality, and to differentiate left from right. In a still this shows as feet that do not quite meet the floor, hands passing through objects they should be gripping, shadows that do not agree with the light source, and reversed or mirrored asymmetric details.
Degraded text. Signage, labels, packaging and clothing text tend to be soft or malformed, and it is typically worse than an image generator at the same nominal pixel dimensions because video encoding is spending its bitrate on motion rather than fine detail.
Pixel dimensions matching Sora 2's supported output sizes: 480x480, 480x854, 854x480, 720x720, 720x1280, 1280x720, 1080x1080, 1080x1920 or 1920x1080, with 720x1280 the API default and portrait framing common because the app feed was portrait-first. Maximum output was 1080p, never 4K. This is weak corroboration rather than evidence, since these are ordinary video dimensions.
Metadata & provenance
Treat metadata as a bonus, not a test. The single most important point for this page: C2PA Content Credentials live in the video container, so they do not survive frame extraction or screenshotting at all. A frame grab is a new file written by whatever tool made it, carrying that tool's tags and nothing from Sora. If you are checking a still, provenance metadata is almost certainly already gone for reasons that have nothing to do with anyone hiding anything. On the original video files, OpenAI stated that all Sora videos embed C2PA metadata. Independent testing published in November 2025, using the official Content Credentials Verify tool and the c2pa-rs command line tool, did not find that to hold uniformly: downloads carrying the visible watermark showed no detectable C2PA, while non-watermarked Pro-tier downloads did. So even for intact video files, a C2PA check that comes back empty is not evidence the file is genuine. SynthID does not apply here on the evidence available. OpenAI announced on 19 May 2026 that it had joined the C2PA steering committee and was embedding Google DeepMind's SynthID invisible watermark, but that announcement covers images generated through ChatGPT, the OpenAI API and Codex, and does not mention Sora. It also postdates the Sora app shutdown by roughly three weeks. We found no evidence that SynthID was ever applied to Sora output, so do not expect a SynthID hit on a Sora frame and do not read its absence as meaning anything. We found no evidence that Sora wrote an EXIF or XMP generator tag naming itself. Separately, C2PA metadata is easy to remove deliberately and is stripped routinely and automatically by most social platforms on upload, so a file that has been through Instagram, X, TikTok or a messaging app will arrive with nothing regardless of origin.
How well WeDetect handles Sora (OpenAI)
Be straightforward here: we have not published a Sora-specific accuracy figure, and this page does not claim one. Our published numbers are roughly 89% overall on modern generators and roughly 85% recall on Midjourney, DALL-E, SDXL and Flux. On Stable Diffusion 3.5 our measured recall is about 62%, which is well below the rest and which we would rather state than bury. Sora frame grabs sit outside that measured set in a way worth naming: our benchmarks are built on native stills, and a frame pulled from a compressed video is a different signal, with encoding artefacts laid over whatever generator traces are present. Video compression can plausibly push results in either direction, and we have not measured which. Detection is probabilistic in every case. It returns a likelihood, not a verdict, and it should be treated as one input alongside the visual checks above and the source and context of the file, not as proof on its own.
Check an image now
Drop the image in below — free, no sign-up, deleted after analysis.
Drag & drop, paste, or click to upload
PNG · JPG · WEBP · HEIC — your image is analyzed privately and deleted
This is an estimate, not proof — the detector can be wrong, producing both false positives and false negatives. Don't rely on it as your only source; verify anything important. wedetect.ai is not liable for decisions made from these results. See our Terms.
Frequently asked questions
Does Sora put a watermark on videos?
Standard downloads from the Sora app carried a visible moving watermark, an animated cloud logo that shifted position and opacity between frames, usually with the account username. Downloads on the $200/month Pro tier came without it. Tools to remove the watermark appeared within about a week of Sora 2's release, and cropping the edges removes it in seconds. So a watermark is reasonable evidence a clip came from Sora, but no watermark tells you nothing either way.
How can I tell if a screenshot is from a Sora video?
Start with the watermark, since a moving cloud logo or a username burned into the frame is the clearest single sign. After that, look for signs it is a video frame rather than a generated still: blocky compression in flat areas, colour smearing at sharp edges, motion blur on moving subjects. Then check the content itself for contact errors, hands passing through objects, shadows disagreeing with the light, and soft or malformed text. Do not expect metadata to help, because frame grabs do not inherit the original file's Content Credentials.
Can you remove the Sora watermark, and does that make a video undetectable?
Yes, removal was widely possible, through dedicated tools, through cropping, or by downloading on the Pro tier where the watermark was never applied. This is exactly why the watermark is useful as positive evidence and useless as negative evidence. Removing it does not remove the underlying visual artefacts, which is why the frame-level checks matter more than the overlay.
Is Sora still available in 2026?
No, not as a consumer product. OpenAI announced the discontinuation on 24 March 2026 and shut down the Sora web and mobile apps on 26 April 2026, with the developer API scheduled to close on 24 September 2026. Sora content still circulates widely as archived downloads, reuploads and screenshots, so images people ask about are typically re-shared material rather than something generated recently in the app.
Other generators: Midjourney · DALL·E 3 · GPT-4o · Stable Diffusion · Flux · Grok · Firefly · Gemini · SDXL · Nano Banana · GPT Image 2 · Flux 2 · SD 3.5 · all generators.