blinktrace · development record
From raw frames to a research system.
A technical timeline of how BlinkTrace moved from raw data processing to held-out evaluation, vision-focused experimentation, and deployment-minded research operations.
01 · milestones
phase 1
complete
Built the preprocessing pipeline
Established the path from raw video to frames, face crops, landmarks, normalized geometry, and sequence manifests. This created the data backbone needed for repeatable experimentation and future streaming-oriented workflows.
phase 2
complete
Established the spatial CNN baseline
Built an appearance-based detector around EfficientNet-B0 face crops and started evaluating it under held-out generation methods. The key lesson: benchmark-style performance means very little without stronger split discipline.
phase 3
complete
Established the temporal GRU baseline
Added landmark-sequence modeling with grouped, source-matched validation. This made it possible to test whether temporal facial dynamics add robust transfer signal beyond still-frame artifacts.
phase 4
in progress
Analyzing fusion and deployment relevance
Fusion analysis shows a nuanced result: combining visual branches can improve hard classification accuracy without necessarily improving ranking strength. That is directly relevant to real-time operating-point design.
phase 5
complete
Introduced orchestration and reporting
Added automation for dataset intake, delta-only preprocessing, validation checks, and report generation. This makes the visual pipeline repeatable and pushes BlinkTrace closer to production-style detection workflows.
current
next
Sharpening the real-time story
The current phase aligns the public presentation with the actual research: rigorous evaluation, deployment-minded experimentation, and a clear direction toward real-time deepfake detection.
02 · where things stand
What BlinkTrace demonstrates today
The project already shows evaluation maturity, vision-focused experimentation, and repeatable research operations — the kind of judgment needed to move from offline experiments toward a real-time detection pipeline.
What comes next
A public repository is coming soon, and the paper will be posted as a preprint on arXiv. For now, this site focuses on the technical throughline and the most meaningful findings.