Seedance is ByteDance's video generation model family, served through the ModelArk platform and integrated into many third-party apps. Since mid-2026, ModelArk has offered an official artifact verification API that checks whether a video carries technical signals embedded during official generation. That API turns "is this Seedance?" from a guessing game into a query with a documented answer — but only if you know what the answer does and does not mean. This guide walks through a defensible order of operations: run the official check first, then corroborate with file evidence and visual review.
1. Run official artifact verification first
The official workflow checks for technical signals associated with ModelArk output. Unlike general-purpose "AI detectors" that estimate a probability from pixel statistics — and are notorious for false positives on filtered or compressed footage — artifact verification looks for provenance signals that the generation pipeline itself embeds. It is more specific than judging whether a clip merely "looks AI-generated," and when it returns a match it usually also returns the exact model name, such as doubao-seedance-2-0.
You can run this check on this site's detector: upload the file or paste a public direct link, and the system creates verification tasks against both official endpoints (Volcengine and BytePlus) in parallel. Results usually arrive within one to two minutes. Three outcomes are possible: official signals detected, no significant signals detected, or unable to determine. How to read each field is covered in the companion note on reading ModelArk verification results.
A match means the video may be an official artifact. A non-match does not prove that Seedance was not involved — re-encoding, cropping and platform compression can remove the signals the API relies on.
2. Inspect the original file and its metadata
Verification works best on the most original file you can obtain. Before or alongside the official check, look at what the file says about itself. Review creation time, encoder, container, resolution, frame rate and signs of repeated transcoding. Useful questions to ask:
- Where did this copy come from? A file downloaded from a messaging app or a social platform has typically been re-compressed at least once. Ask the source for the original export when the stakes justify it.
- Does the timeline make sense? Compare the file's creation and modification timestamps against when the event supposedly happened and when the clip was first published.
- What wrote the file? Encoder and writing-library tags sometimes reveal an editing tool or a platform pipeline rather than a camera. A "camera original" that was written by a video editor deserves more questions.
- Has it been through multiple generations? Stacked compression artifacts, resolution that does not match any common camera preset, or a container/codec combination typical of export pipelines all suggest distance from the original.
Metadata is corroborating evidence, not proof: it is trivial to strip and possible to forge. Treat it as one line of evidence that either supports or contradicts the claimed origin story.
3. Look for cross-frame continuity
Modern video models can produce excellent individual frames, so single-frame "AI tells" are increasingly unreliable. The weak point is continuity: keeping every object structurally consistent across motion, occlusion and lighting changes. Scrub the clip frame by frame around the moments where consistency is hardest to maintain.
| Signal | What to inspect |
|---|---|
| Hands and contact | Structural drift after occlusion — fingers that merge, split or change count when a hand passes behind an object and reappears |
| Text and logos | Strokes, spelling or letterforms that change shape across frames while the camera moves |
| Motion physics | Unexplained speed changes, objects that decelerate without cause, collisions that do not transfer momentum |
| Reflections and shadows | A subject and its reflection moving differently, or shadows that do not track the light source |
| Background crowds | People or vehicles far from the focal subject that pop in and out, repeat or deform |
None of these clues is conclusive on its own — real footage can contain compression weirdness, and a clean clip can still be generated. Visual review tells you where to look harder; the official check tells you what the generation pipeline recorded.
4. Use a review workflow that preserves uncertainty
If you review content professionally — as a journalist, moderator or platform operator — the goal is a repeatable process whose conclusion you can defend later, not a hunch:
- Keep the original file and the source URL. Never work only from a re-uploaded copy.
- Run official verification and record the result, including the model name if one is returned.
- Compare file metadata with the claimed publishing chain.
- Review the important segments frame by frame using the continuity checklist above.
- Report matched, not matched or unable to determine — with the evidence for each — rather than forcing a binary answer.
Know the limits before you conclude
The official artifact verification API covers the Seedance model family only: Seedance 2.0 and later video models, and Seedream 5.0 and later image models. It does not detect Sora, Kling, Runway, Veo or face-swap deepfakes. It verifies artifacts generated within roughly the last 30 days, and heavy re-encoding can remove the signals it depends on. In practice that means a detected result is strong evidence, while a not detected result is weak evidence — it should lower, not eliminate, your suspicion, and it says nothing about other AI models. Automated analysis supports review decisions; it is not a legal or forensic determination.