Open source deepfake detection: which codebases publish weights, and under what licence

This page belongs to the deepfake detection hub. It answers a practical question that the scores on the cross-dataset generalisation page leave open: if a detector's number is worth checking, where can its trained model be downloaded, and what may be done with it?

01Which repositories publish weights?

The table lists what each repository's README, licence file and release listing said on 10 November 2026. "Not stated" means the repository does not say; it does not mean the weights are free to use.

Each row comes from that repository's own files: the DeepfakeBench README and its releases, the SBI README, the GenConViT README and its Hugging Face model card, and the other repositories listed in the sources below.

02What does DeepfakeBench release?

DeepfakeBench (Yan et al., NeurIPS 2023) is the most complete source because it is a framework, not a single method. Its README states that "we have provided the pre-trained weights for each detector" and links to release v1.0.1, titled "Release the pre-trained weights for all detectors". That release holds 13 files, named after Meso4, MesoInception, CNN-Aug, Xception, EfficientNet-B4, Capsule, CORE, F3Net, FFD, RECCE, SPSL, SRM and UCF. No file in it carries the name of DSP-FWA or Face X-ray, two of the 15 detectors in the original paper; Face X-ray has a separate release of the auxiliary files it needs.

Three other releases hold supporting files. Release v1.0.0 contains the ImageNet backbone weights in pretrained.zip, which the README says "are from ImageNet", plus the landmark detector and the supplementary material. Release v1.0.3 adds a 3D ResNet-50 weight file for the video detectors I3D, FTCN and AltFreezing. The README now lists 36 detectors, 28 image and 8 video, so a release made for the first 15 does not cover the rest. The framework and its reported scores are described on the benchmark's own page.

03What do the licences allow?

The CC BY-NC 4.0 licence on DeepfakeBench has one operative restriction for readers who want to reuse the weights: the NonCommercial condition, which Creative Commons words as "You may not use the material for commercial purposes". Research use, with attribution, is allowed. The repository's LICENSE file states the terms for "the Software source code" and names third-party components as under other licences, so the licence text does not mention the weight files by name.

The others differ. MIT (GenConViT code and the DFDC solution) and Apache 2.0 (MesoNet) are permissive licences that allow commercial reuse of the code with attribution. The SBI repository's licence file is titled "ACADEMIC OR NON-PROFIT ORGANIZATION NONCOMMERCIAL RESEARCH USE ONLY", and its README adds that for commercial use "A license agreement is required." LAA-Net uses an academic licence that grants use for "academic or scientific research purposes only". FTCN lists a weight file and no licence at all, which leaves the terms of use unclear rather than open.

04Why can the weight licence differ from the code licence?

A repository can hold two separate works: source code and a trained model. GenConViT shows the split in one project. Its GitHub README ends with "MIT License", while the Hugging Face model card that hosts the two weight files states "license: cc-by-nc-4.0". A reader who takes the code under MIT and the weights from the model card is bound by the noncommercial terms for the weights.

Training data adds a third layer. A detector's weights were produced by training on datasets, and datasets have their own terms: FaceForensics++, for example, is distributed by a Google form under the FaceForensics Terms of Use. Whether those terms pass through to a trained model is a legal question that none of these READMEs answers. The datasets page records the access terms of each public set, which is the starting point for that check.

05Do the published weights reproduce the paper's numbers?

Only where the weights match the paper's training conditions. The SBI README publishes two EfficientNet-B4 files, one trained on FaceForensics++ raw and one on its c23 release, and says that "Using the provided pretrained model, our cross-dataset results are reproduced as follows". DeepfakeBench trains every detector on FaceForensics++ c23, so its scores and the SBI raw weights come from different training conditions. The reproduction guide explains how to line up the protocol before comparing two numbers.

Hosting matters too. Weights on a GitHub release or in the repository sit beside the code and its history. Weights behind a Google Drive, Dropbox or Hugging Face link depend on a third party continuing to host them, and a README can outlive its links. The DFDC solution keeps its seven files in one GitHub release and fetches them with a script, download_weights.sh, so the list of files and the files themselves stay together.

06How to check a repository before using its weights

Five checks, in order, catch most problems.

1. Find the licence file and read what it covers: code, weights, both, or neither.
2. Find the licence for the weights where they are hosted, such as the model card.
3. Note which training set and compression level the weights were trained on.
4. Check where the files are hosted and whether the link still works.
5. Compare the training conditions with the paper before quoting a score from the repository, as the SBI page and the GenConViT page do for their methods.