# Fully open-source path for the Orislop AV detector

## Short answer

Yes, the Orislop AV detector can be open-sourced, but code, model weights, and
training data have separate licenses. The cleanest route is:

1. publish the Orislop architecture and training code under Apache-2.0 or MIT;
2. train new weights from scratch;
3. use only approved, provenance-tracked data such as the AMI CC BY 4.0 corpus;
4. publish the model card, data-source ledger, attribution, metrics, limitations,
   and weight license together.

This kit does not import Wav2Lip or AV-HuBERT checkpoints.

## Relevant existing projects

| Project | What it is useful for | Code license / restriction | Clean default for Orislop? |
|---|---|---|---|
| Orislop AV Joint | Multitask lip-sync, spoof, visual and joint detector | Your chosen project license; weights start from scratch | Yes, after data and dependency review |
| SyncNet | AV synchronization and active-speaker architecture | MIT repository | Architecture/reference is promising; verify any downloaded checkpoint separately |
| TalkNet-ASD | Active-speaker detection | MIT repository | Code is usable; pretrained-weight training-data provenance needs separate review |
| LatentSync | Lip-sync generation, not a detector | Apache-2.0 repository | Not needed for the detector |
| Wav2Lip | Lip-sync generation and expert discriminator | Personal/research/noncommercial | No for a commercial/open release |
| AV-HuBERT | AV speech representation | Custom Meta license | No as the default fully open dependency |

An open repository does not automatically make every checkpoint or dataset in
that repository suitable for commercial use or weight redistribution.

## Dataset decision

- **Enabled:** AMI Meeting Corpus, official close-up video and individual headset
  audio downloads, CC BY 4.0, with published ethics/consent documentation.
  File-level provenance is recorded.
- **Visible but disabled:** CREMA-D. It is open, but ODbL/DbCL obligations need a
  project-specific review before combining or redistributing it.
- **Other reviewed options:** Audio-Visual Lombard is CC BY 4.0 but remains a
  manual source; GRID is disabled until its file-level license is confirmed.
- **Excluded from fully-open mode:** LAV-DF, FakeAVCeleb, AV-Deepfake1M, and DFDC. Their
  request terms, noncommercial components, unclear redistribution terms, or
  separate access agreements prevent automatic inclusion.

## What 50 GB of AMI does and does not solve

AMI is authentic synchronized speech and is a strong base for representation
learning, active-speaker learning, genuine negatives, compression robustness,
and controlled synchronization-offset augmentation. It is not a diverse
collection of modern synthetic generators. A production detector still needs
rights-cleared manipulated examples, legitimate dubbing/delay negatives, and
held-out generator families. Never report an AMI-only model as a general
deepfake detector.

## Publication checklist

- Keep `provenance.jsonl` and the AMI attribution with the release.
- Document every transformation and synthetic augmentation.
- Do not redistribute source videos inside the model repository.
- Keep source/session groups in only one train/validation/test split.
- Release weights as experimental until independent evaluation and shadow-mode
  gates pass.
- Have counsel review privacy/publicity, trademark, and weight-license questions
  before a commercial launch. This kit records evidence; it does not provide a
  legal opinion.