Have you ever wondered why GPTbot or other crawlers download your images in different resolutions?
This serves two purposes:
- Forensic AI (for businesses, governments, and agencies)
- Custom vision models
How does it work in both cases?
Give a model the tiny thumbnail (150×150) and ask it to read/OCR/”see” the content. Give another model the full-resolution image (768×1665 or larger) as ground truth. Compare outputs—train the small-image model to be as accurate as the large-image model
Teacher model: receives the high-res image, produces accurate OCR
Student model: receives the low-res thumbnail, attempts to match the teacher’s output
Training signal: difference between teacher and student outputs
The result would be a vision model that excels at extracting text and much more from degraded images—thumbnails, screenshots, blurry images, compressed social media content.
A third AI or a human in the loop (depending on the importance of learning success) then evaluates the output of the student model against the teacher model. The smaller the loss or BpB relative to the correct prediction or baseline of the teacher model, the better the trained student model.
The student model then becomes the new teacher model.
Conclusion: After training, both Vision models (Forensic/Commercial) can:
- Verify metadata despite poor resolution (e.g., time/battery level in screenshots vs. metadata)
- Accurately read and reproduce even low-resolution text and images
- Capture, summarize, collect, and link content across the web (token tracing with text or visual markers for images)
- Recognize and identify faces and landscapes even in thumbnails
- Extract steganography not only from high-resolution images
- Correctly extrapolate minute details from low-resolution images
That’s why they’re downloading every resolution—not just hoarding data, but executing a specific training strategy.