Super-Resolution Hallucination, Playable

EP 01

The pixel shredder — watch 98% of the information die

Type any character. It gets rendered at 32×32 (1,024 pixels), then downscaled by averaging blocks — exactly what a cheap camera or a small thumbnail does. At 2× you can still read it. At 8× the entire letter lives in 16 numbers. Everything else is gone. Not hidden. Gone.

Character Downscale

The law behind everything below: downscaling is a many-to-one function — thousands of different originals map to the same small image. Inverting it is not "sharpening", it's answering a question with many correct answers. Keep that in mind for every "enhance" you ever see.

EP 02

Two ways back up — honest blur vs confident invention

Bicubic upscaling only interpolates the pixels it has: blurry, but it never claims to know more than it does. The "AI" upscaler here has a prior — it knows the world contains letters and digits — so it finds the character whose downscaled version best explains the observation, and redraws it razor-sharp. Then press the doodle button: feed it something that is not in its prior, and watch it confidently draw a letter anyway.

Character Sensor noise 0.15

This is really how it works: Real-ESRGAN, GFPGAN and diffusion upscalers do a continuous version of this — they learned "what sharp images look like" and pick the plausible one that matches your blurry input. When your input isn't in their world (your grandmother's actual face), they still answer with something from their world.

EP 03

Sharp is not accurate — the error overlay doesn't lie

Which reconstruction is actually closer to the truth? Your eye votes for the sharp one — so let's measure instead. Red marks every pixel where the reconstruction disagrees with the ground truth, and PSNR is computed live. Try the easy preset: AI wins big. Then try the trap: when noise makes the AI guess the wrong letter, it scores worse than honest blur — while still looking better.

Why benchmarks disagree with your eyes: the field measures both distortion (PSNR — am I correct?) and perceptual quality (does it look real?), and there is a proven trade-off between them. AI upscalers deliberately sacrifice correctness for beauty. For wallpaper, great deal. For evidence, terrible one.

EP 04

Two originals, one thumbnail — the identity test

Take two different characters and shrink both 8×. Their small versions become nearly identical — the distance between them is measured live below. Now upscale both with the AI. With no noise it can still split hairs; add realistic sensor noise and both reconstruct as the same character. One of the two identities was destroyed in the shrink, and no algorithm — present or future — can tell which pixels held it.

Pair Sensor noise 0.30

The information-theory heart of it: once two originals map to the same observation, distinguishing them is mathematically impossible — the bits are destroyed. Everything an upscaler "recovers" past that point is prior, not evidence. This is the precise sense in which enhanced detail is fiction.

EP 05

The license plate — why “enhance” is inadmissible

A random 5-character plate is photographed at low resolution, then "enhanced" by the AI. The output is perfectly sharp — every character crisp and confident. Green boxes: it invented the right character. Red boxes: it invented the wrong one, with identical confidence and identical sharpness. Run it a few times at 8× and watch the running accuracy — would you convict on this?

Downscale Sensor noise 0.25

The plate uses easily-confused characters (0/O/D/Q, 8/B, E/F, 1/I) — exactly the ones that collide at low resolution.

The forensics warning: AI-enhanced faces, plates and documents look like recovered truth but are statistical invention — which is why courts have begun rejecting them, and why Samsung's "moon photos" turned out to be textures pasted from a prior. ✓ Good case: making a vacation photo prettier. ✗ Bad case: any pixel someone's fate depends on.