Hallucination

Algorithms invent details that were never filmed — this is the core problem with modern upscaling and frame interpolation. A hallucination occurs when AI models, trained on millions of images, simply add plausible-looking pixels to incomplete or ambiguous input. The computer "sees" something that the cinematographer didn't have in front of the lens.

In practice, this happens constantly: you upscale 2K material to 4K, and suddenly houses in the background have windows that never existed. With frame interpolation — meaning the calculation of intermediate frames for slow motion — motion sequences are invented that are biomechanically incorrect. A hand moves across a table, and the algorithm fabricates fingers in impossible positions from the blurry input. This isn't just blur or noise — it's active invention.

The reason lies in the statistical nature of neural networks: they learn probability distributions, not exact representations. If the information is too low-resolution or ambiguous for the model, it chooses the "most probable" solution — and that is often plausible enough not to be immediately noticed as an error. It becomes particularly problematic with extreme upscaling (4:1 factor and higher), with fast movements, or with structures that the training dataset has rarely encountered. A detail on a face, lettering on a sign, reflections in glass — all of this can be hallucinated.

On set and in the edit, you need an eye for it: don't blindly trust automatic upscaling tools. Compare the original and the result side-by-side, especially in details that are not narratively important (invented artifacts are less noticeable there). Some projects fare better with native shooting in the higher resolution than with downstream processing. With frame interpolation, it's similar — no matter how clever the motion calculation, without the original material, it remains speculative. For critical shots where credibility counts, there's no way around classic interpolation or re-shooting.