Orislop AV Joint
Orislop AV Joint is an experimental, multitask audio-visual forensics training
kit. It contains a T4-oriented Google Colab workflow, a PyTorch model trained
from scratch, bounded AV preprocessing, evaluation/calibration/export commands,
and a rights-aware downloader for approved open audiovisual sources.
Repository status
This repository contains training code, not trained or production-approved
weights. Do not present it as a working universal deepfake detector until a
checkpoint has passed the documented data, evaluation, calibration, fusion, and
shadow-mode gates.
The exported artifact remains unpromoted by default. Standalone AV output is not
authorized to hide or skip content.
Colab quick start
- Download this repository as a ZIP.
- Open
Orislop_AV_Joint_Open_Data_Trainer.ipynb in Google Colab.
- Select a T4 GPU.
- Run the notebook in order.
- To acquire AMI data, read the linked AMI license and consent evidence, then
change the two explicit acceptance switches in the acquisition cell.
The acquisition tool is not a general web scraper. It only follows the official
AMI corpus host allowlist, caps completed AV media at 50 decimal GB, resumes
partial files, rejects off-host redirects, and records SHA-256 provenance.
On Windows, tools/run_local_data_pipeline.ps1 provides plan, discovery,
download, manifest-build, and full modes. Discovery accepts no license and
downloads no media. Full mode requires the explicit -AcceptAmiLicense switch,
checks target-drive capacity, and defaults to D:\OrislopAVData.
If the standard Hugging Face CLI reports a Windows certificate-chain error,
tools/hf_native_tls.py runs the same CLI through the operating system trust
store. tools/hf_native_tls_token_login.py provides a hidden-input token login
without putting the credential in a command line.
The paired manifest generator reads AMI's official meeting-specific
camera/headset mapping. It produces both synchronized examples and controlled
audio offsets. Controlled offsets are labeled as legitimate delay
(sync_mismatch=1, joint_forgery=0), not synthetic-media fraud.
Open-source and data boundary
Code, trained weights, and datasets have separate licenses. Read:
OPEN_SOURCE_PATH.mddata_sources.jsonAMI_ATTRIBUTION.mdDATASET_LABELING_GUIDE.md
No project-level software license has been asserted in this model card. Add a
reviewed LICENSE before claiming that the repository or future weights are
fully open source.
AMI-only training is useful for authentic AV representation and synchronization
experiments, but it is not sufficient for a general deepfake detector. Such a
release also needs rights-cleared manipulated examples, legitimate dubbing and
delay negatives, held-out generator families, speaker/source-disjoint splits,
calibration, independent evaluation, and reviewed shadow decisions.
Safety and limitations
- Predictions are probabilistic signals, not proof of deception.
- Ordinary dubbing, translation, editing delay, network lag, and accessibility
audio must not be labeled as forgery solely because AV timing differs.
- The detector must abstain when face coverage, speech, motion, SNR, or
uncertainty gates fail.
- Do not use it to make employment, credit, housing, identity, or legal
decisions.
- Do not redistribute source media unless its license and consent terms permit
that exact use.
- Treat synchronization mismatch as a forensic signal, never as proof that a
person lied or that a video is malicious.
