ByteDance has disclosed the scale and broad sources of the data behind Seedance 2.5. For an illustrator, photographer or filmmaker trying to trace one piece of work, however, the public record still stops short of a usable answer.
When ByteDance introduced Seedance 2.5 on 31 July 2026, the product headlines were easy to grasp: videos of up to 30 seconds, more reference material and finer control over edits. A second publication dated the same day deserves at least as much attention from people who make the images and footage that circulate online.
The company’s Public Summary of Training Data Content for Seedance 2.5 says the model was trained on more than one billion images, more than one million hours of video and more than one million hours of audio. It describes a mixture of publicly available, licensed and synthetic data. It also identifies Bytespider as the crawler used to collect publicly accessible images, video and audio through June 2026.
A corresponding summary for Seedance 2.0, updated on the same date, reports the same broad scale: more than one billion images and more than one million hours each of video and audio. These documents make the size of the training operation visible. They do not make it possible for a creator to search for a photograph, illustration or clip and learn whether that particular work was included.
For an individual creator, the aggregate figures do not answer the work-level question.
The summaries explain scale, but not provenance at the level creators need
Both summaries contain information worth having. ByteDance says it used public, licensed and synthetic material. It says commercial licensing agreements cover image, video and audio content. It also says that interactions with the models, and interactions with ByteDance’s other products, were not used to train them.
For web-crawled material, the company says Bytespider is designed to follow robots.txt instructions and avoid paywalls and password-protected content. That describes how ByteDance says Bytespider was configured during data collection. It does not identify whether a particular publicly accessible file entered the corpus.
The missing detail is especially visible in the sections intended to identify sources. The Seedance 2.5 summary does not name the principal public datasets. In the field for the most relevant domains crawled, it gives a description of subject matter—resources spanning natural scenes, objects, people and diagrams—rather than domain names. Licensing counterparties are not identified either.
No single-work search or verification process is described in the public summaries. That is not proof that ByteDance has no internal provenance records or private claims process. It means only that the material available to the public does not show a creator how to check one work.
The studio agreement reveals a participation gap
On 17 August 2026, the Motion Picture Association and ByteDance announced a memorandum of understanding covering intellectual-property safeguards for Seedance, Seedream and the products through which they are offered. The announcement followed objections raised by the MPA after the launch of Seedance 2.0 and described continuing collaboration on stronger guardrails.
The announcement says the parties held constructive conversations after the MPA’s February cease-and-desist letter and that newer model launches reflected advances in IP protection. It does not publish the memorandum or describe a work-level search, a claims route for independent creators, licensing payments or compensation. It is an agreement about guardrails, not a disclosure of the models’ training data.
Output safeguards and training-data provenance are related, but they are not the same problem. A model may block a request involving a well-known film character without giving a photographer any way to ask whether her original image entered the training set. Equally, the presence of a work in a dataset would not, by itself, settle every legal question about a particular generated output.
The public record shows a negotiation channel for the MPA. It does not identify an equivalent route for independent animators, photographers, small production houses or community archives. That is the participation gap at issue here.
In the Global South, the issue extends beyond individual copyright
A June 2026 issue brief prepared by IT for Change for the AI, Culture and Intellectual Property subgroup of UNESCO’s Global Civil Society Organizations and Academic Network on AI Ethics and Policy places this debate in a wider setting. It argues that cultural resources, public knowledge infrastructures and social data from the Global South are frequently drawn into AI development while economic value is concentrated among a small number of firms. UNESCO notes that the subgroup’s work does not necessarily represent the organisation or its member states.
This raises a question that the ByteDance summaries do not address: is technical accessibility enough when material is culturally sensitive or collectively held? A local creator may post work for education, documentation or discovery without intending it for model training.
ByteDance’s summaries say no geographic region was intentionally excluded and that the video data contains a range of global languages. That broad reach makes regional accountability more important, not less. Yet the summaries do not provide a geographic or linguistic breakdown that would let communities examine how their material is represented, licensed or filtered.
Many independent creators will also benefit from easier access to capable video tools. The same person may be a model user and a rightsholder: a photographer can generate a campaign on Monday and spend Tuesday trying to trace how an older image travelled online. A workable governance system has to recognise both roles.
Access services address a different question. reAPI’s Seedance 2.0 page documents an API route, while ClipDance’s Seedance 2.5 page offers a browser workflow. Neither page identifies ByteDance’s training sources.
Useful transparency begins with a question a creator can actually ask
Creators do not need a public download of the training set, which could expose personal information, licensed assets and the works of other people. They do need a way to move from a general disclosure to a specific, reviewable claim.
One approach would allow a rightsholder to submit a file fingerprint, source URL or other evidence and receive a bounded answer about whether a match exists. Where a direct answer would reveal protected information, the process could use an independent auditor or a confidential review. The response should explain what evidence was checked, what the result means and how it can be appealed.
Source reporting could also become more concrete without publishing the corpus itself. Providers could name principal public datasets and domain categories, give collection windows, distinguish crawled material from licensed material, and report how rights reservations are honoured. Aggregate figures on requests, matches, removals and response times would show whether the process works beyond a policy page.
Collective representation is just as important. Negotiations should not be limited to the best-resourced entertainment companies. Photographers’ groups, independent film associations, archives, Indigenous communities and cultural organisations need channels that recognise both individual rights and collective custodianship. Those channels must be usable across languages and jurisdictions, including places where a creator cannot afford specialist counsel.
ByteDance’s summaries disclose the scale, crawler and broad data categories behind the models. They still leave creators without a public way to check a known work. That is the specific transparency gap the company has yet to close.