Deepfake dataset license terms: what research-only means in practice

This page belongs to the deepfake detection hub and covers licences, the one part of a dataset card that the public deepfake datasets compared page can only summarise in a column. It reads the terms clause by clause. Whether the people shown agreed to appear is a separate question, answered on the consent in deepfake datasets page.

01What does a research-only clause allow?

A research-only clause gives a researcher permission to use a dataset for research and withholds the same permission for commercial ends. The FaceForensics terms of use, hosted by the Technical University of Munich, put it in the first clause: "Researcher shall use the Database only for non-commercial research and educational purposes." The Celeb-DF terms open the same way, with the dataset "for non-commercial research purposes only".

Neither text defines "non-commercial". The nearest published definition is the one in the Creative Commons legal code for CC BY-NC 4.0, where NonCommercial "means not primarily intended for or directed towards commercial advantage or monetary compensation" (Creative Commons legal code). That sentence is a licence definition, not an interpretation of the two form-based terms, so it shows how the phrase is usually read without settling the edge cases: a company research team, a free product funded by advertising, a detector offered to paying customers after training on the data. This page reports what the texts say and does not give legal advice.

02Which terms govern each public dataset?

Five sources cover the sets that most benchmark papers use. The table gives how access works, which document governs, and the use limit in that document's own words.

The FaceForensics and Celeb-DF rows come from the terms documents and the FaceForensics++ page and Celeb-DF page describe the sets themselves. The DFDC row is shorter because the Meta page says what the data is and how to sign up but names no licence in its text. It states that the data "was created by Facebook with paid actors who entered into an agreement to the use and manipulation of their likenesses", which is about the people shown, not about the person downloading. The DFDC page holds the paper's counts.

03How do the FaceForensics and Celeb-DF terms differ?

Both texts are short lists of numbered clauses, seven for FaceForensics and six for Celeb-DF, and they overlap on the main point. They diverge on sharing, liability and the end of access.

The sharing row matters most for working practice. Under the FaceForensics terms a lab can give a colleague access if the colleague accepts the same terms. Under the Celeb-DF terms the user agrees not to "publish or distribute any portion of the Celeb-DF dataset or any derived data". On a plain reading, derived data covers cropped faces and extracted frames, so a preprocessed copy placed on a shared drive for people outside the agreement is outside what the terms allow.

04Does CC BY-NC on a benchmark repository cover the datasets it runs on?

No. The DeepfakeBench README states that the repository is licensed by The Chinese University of Hong Kong, Shenzhen under CC BY-NC 4.0, and the LICENSE file applies that licence to the Software "except for the third-party components or materials listed as follows which are subject to different license". The repository holds code. The datasets it evaluates sit in their own repositories, each with its own terms: the README's dataset table links FaceForensics++ and Celeb-DF to the repositories described above.

That table has a column headed "Rights Cleared". It reads NO for FaceForensics++, FaceShifter, Celeb-DF v1, Celeb-DF v2 and UADFV, and YES for DeepfakeDetection, the DFDC preview, the full DFDC and DeeperForensics-1.0. The README does not define the column. The YES rows are the sets filmed with recruited actors: the DFDC paper reports "3,426 paid actors" who agreed to be modified (arXiv 2006.07397), and the same table lists 28 subjects for DeepfakeDetection and 100 for DeeperForensics-1.0. The DeepfakeBench page covers the framework; the detector numbers trained on FaceForensics++ at c23 sit in the published detector scores.

05Do research-only terms reach trained models?

The texts read for this page are silent on model weights, with one pointer. The Celeb-DF clause covers "any derived data from the dataset for any purpose", and whether a trained detector counts as derived data is a question the text does not settle. The FaceForensics terms limit use of "the Database" and say nothing on models.

Face recognition offers a clear example of a project that answered the question in writing. The InsightFace repository releases its code under the MIT licence but states that its training data and the models trained on it are for non-commercial research only; the InsightFace page records the wording. A deepfake detection team that plans to release weights can borrow the habit: state the terms of every training set next to the weights.

06What happens when access is withdrawn?

Each source handles it differently. The FaceForensics terms keep the right to "terminate Researcher's access to the Database at any time". Celeb-DF's README says that anyone who is uncomfortable about their identity shown in the dataset can write to the authors, who "will remove corresponding information from the dataset". The DFDC page rests on the agreements the paid actors signed. None of the texts sets a deadline or a notice period. A lab that publishes a result on a dataset should therefore keep a note of the terms version it read and the date, because access can end while the paper stays in print.

07How to check a dataset's terms before you train

1. Find the document the README names: the terms of use, the form text or the licence file.
2. Read the use limit and the sharing clause, and note who else is bound, such as an employer or an institution.
3. Check how the text treats derived data: crops, frames, features and weights.
4. Check whether the form-gated copy and the public licence agree, as for a benchmark that redistributes a set.
5. Write down the date you read the terms. This page read each source on 10 November 2026.