"Training Data Maxxxing Month" starts with a simple question:
Does the model care where the video frames came from?
We just released 'Temporal Fidelity', a controlled benchmark with matched source clips in:
• native 60p • clean 24p • synthetic 30p derived from the 24p version
The goal is not just to ask whether 60p is better. (it is)
We want to know:
• Does temporal contamination hurt model performance? • Can models detect temporal junk? • Can temporal provenance become a useful filtering signal for scraped video datasets? • Does a file labeled 30 fps actually contain 30 unique captured moments per second?
If you test it, find something weird, or build a detector with it, we'd love to see the results.
We’re releasing one controlled video-data test every week in October, free on Hugging Face.
The goal is simple: isolate one variable at a time, then stack them.
Oct 5: 24p vs 30p vs 60p Oct 12: SDR vs HDR Oct 19: 1080p vs 4K vs 8K Oct 26: 8K + 60p + HDR
Same scenes. Controlled comparisons. Real-world footage.
If you train or evaluate video models, we want to see what changes: evals, ablations, side-by-side generations, temporal consistency, training efficiency, or any measurable model improvement.
The point isn’t to assume higher-quality source data is better.
It’s to test it.
Week 4 combines everything into the full Overlai "Gold Standard" stack.
I wanted to introduce myself and my company @Overlaiapp. We are a collective of filmmakers, photographers, and AI engineers working on high resolution (8K+) training data.
We plan to share a lot of our datasets with the community and are kicking things off with two curated datasets:
🎥 Oversampled: Every clip is captured in stunning 8K resolution, delivering rich detail ideal for fine tuning scenic landscapes and ocean dynamics.
📸 Variance: Includes close-up details, slow-motion footage of crashing waves, sweeping landscapes, and wildlife shots.
📋 Detailed Metadata: Every clip is paired with structured metadata, including creative descriptions, precise camera movements, lens information, field of view calculations, and shot settings, ensuring AI models can fully understand and replicate real-world cinematography with accuracy.
⚙️ Consistency: Re-thinking training data at the point of capture by "overshooting" a subject, enabling models to learn more nuanced relationships and views across scenes.
🌅 Light: Shot during early morning and sunset light for optimal color contrast and dynamic range, maximizing visual quality for color and lighting-sensitive tasks.
🔍 Curation: Curated specifically for machine learning, providing clean, high-quality data for next generation model training.
