How Mosaic Censorship Actually Works, Technically

Mosaic censorship replaces a region of each frame with a grid of solid blocks, each block painted the average colour of the pixels it covered. Averaging is a one-way operation applied to the master before encoding, so the detail inside a block is discarded at source and no later process can recover it.

Last updated Mon Aug 03 2026 00:00:00 GMT+0000 (Coordinated Universal Time)

A mosaic divides a region of the frame into a grid and replaces every pixel in each cell with a single value — usually the average of the pixels that cell covered. Many values go in, one value comes out. That is the entire operation, and it is the reason the result cannot be undone.

Understanding it at that level answers a lot of adjacent questions at once: why block size is the only meaningful variable, why the censorship is baked into the file rather than applied by the site you are watching on, and why "AI uncensoring" is a category error rather than a difficult problem.

What does a mosaic actually do to the pixels?

It performs block averaging: partition the target region into a grid of cells, compute one colour per cell from the pixels inside it, then paint the whole cell that colour.

Take a 16×16 cell in an ordinary video frame. Before the operation it holds 256 pixels, each with its own colour — 256 independent values describing edges, texture and shading. After the operation it holds one value, repeated 256 times. The 255 degrees of freedom that described the detail are gone from the file. They were not moved, hidden or encrypted. They were averaged away.

This is what mathematicians call a many-to-one mapping. An enormous number of different source blocks produce the exact same average, so given the average there is no way to determine which source produced it. That is not a weakness in the implementation; it is the defining property of the operation, and it is why the technique was chosen.

Why is block size the only real variable?

Because everything else about a mosaic is fixed by the method. The only choice that changes how much information survives is how large each cell is.

Larger cells average more pixels together and destroy more. Smaller cells preserve more structure — enough, at the small end, for shape to remain legible, which is exactly what review standards exist to prevent. Coarseness is therefore what the standards regulate; Japan's review bodies name mosaic density among their review criteria.

How they express that requirement is not public. The bodies do not publish their expression rules — OCCN, one of the three with a live public site, states that they are disclosed to members only — so whether coarseness is defined as a proportion of frame dimensions or in absolute pixels, and at what value, cannot be sourced. Treat any specific figure you encounter as unattributed. What is worth knowing regardless is that a fixed pixel size would mean something completely different at SD and at 4K, so the two framings are not interchangeable.

Sources: OCCN (occn.or.jp) review and concept pages; JCRC (jcrc.or.jp) site structure; checked 2026-08-03.

Treatment What it does What survives Reversible?
Mosaic / pixelation Averages each grid cell to one value Low-frequency colour and rough position No — inputs are collapsed
Gaussian blur Weighted average across a moving window More low-frequency structure than mosaic No, though it degrades less
Solid bar or fill Overwrites the region with a constant Nothing about the region No
Crop or reframe Removes the region from the frame Nothing No
Overlay on a player Draws on top at playback time Everything — the file is intact Yes, trivially

Only the last row is a real security question, and it is the one method the industry does not use.

At what point in production is the mosaic applied?

During post-production, to the master, before the title is submitted for review — not at playback, not by the platform, and not by the encoder that prepares a streaming rendition.

That ordering is deliberate and it is the whole design. Because censorship is applied upstream of everything else, every downstream artefact inherits it: the review copy, the retail file, each streaming bitrate ladder rung, every screen capture a viewer takes. There is no point in the chain where an uncensored version exists that someone could get at by intercepting a stream or requesting a different quality level.

It also means the platform hosting a title has never held the uncensored footage. Asking a streaming site for an uncensored copy is asking for something it was never sent.

Why can't the original detail be recovered?

Because the file no longer contains it. Recovery would require inverting an operation that has no inverse.

Compare it with things that are reversible. A compressed file can be decompressed because compression stores a recipe for reconstruction. An encrypted file can be decrypted because the content is intact under a transformation with a key. A mosaic has neither: no recipe, no key, no transformed copy of the data. There is a single number where a texture used to be.

The practical test is simple. If you took a thousand different source blocks that all average to the same grey, mosaic them, and hand someone the results, they are identical files. No process, however sophisticated, can tell you which of the thousand it started from — because the information that distinguished them is not present in the input.

What does AI "restoration" actually produce?

Plausible invention. A generative model trained on uncensored imagery can produce a patch that is consistent with the surrounding pixels and with the block averages, and it will look convincing. What it cannot do is be correct, because correctness is not determined by anything in the file.

The distinction that matters is between reconstruction and hallucination:

  • Reconstruction derives output from information present in the input. The answer is verifiable against the source.
  • Hallucination generates output consistent with the input's constraints. Many different outputs satisfy the same constraints equally well, and nothing distinguishes them.

Mosaic removal can only ever be the second. The model is not uncovering what was there; it is drawing something that would have averaged to the same blocks. Presenting that output as the original is a factual claim the process cannot support — which is a separate problem from any legal or ethical one.

This is the same reliability question that runs through AI-generated content generally: a system that always produces confident output tells you nothing by producing output.

How does mosaic interact with video compression?

Favourably, which surprises people. Mosaic regions are cheap to encode.

Video codecs spend bits on high-frequency detail and on change between frames. A mosaic region is the opposite of both: large flat areas of constant colour, with edges that fall on a regular grid. Block-based codecs represent that efficiently. The censored region typically consumes fewer bits than the detailed footage it replaced, so the rest of the frame gets more of the budget.

One visible side effect is that mosaic edges stay crisp even at low bitrates where the surrounding image goes soft. The blocks are not a compression artefact, though they are often mistaken for one — a real compression artefact shifts with motion and bitrate, while a mosaic grid stays locked to the region it covers.

Does the mosaic interfere with identifying a performer from a frame?

No, because the mosaic never covers the face. Every step of face-based matching operates on pixels the censorship did not touch.

Detection finds the facial region and its landmarks, alignment normalises pose using those landmarks, and an encoder converts the aligned crop into a vector — 512 numbers in our case. None of that machinery looks anywhere near the censored region. A heavily censored title and an uncensored one are equally legible to a face encoder.

What actually limits identification is which sources were indexed in the first place. Our index covers 241,792 faces across 106 sites, but 2,246 of the 2,333 named performers have their representative image from a single site (our index, 2026-08 snapshot) — a source concentration that constrains results far more than any censorship standard.

Related questions

Frequently asked

Is a mosaic reversible if you know the block size?
No. Each block stores one value where there were many, so the original pixels cannot be derived from it. Knowing the grid geometry tells you where the information was destroyed, not what it was.
Do AI models restore censored video?
Generative models synthesise plausible detail consistent with the block averages. That output is invented, not recovered, and there is no way to check it against an original that the file no longer contains.
Does the mosaic make video files harder to compress?
The opposite. A mosaic region is large flat areas with sharp edges, which most codecs encode very cheaply, so the censored area typically costs fewer bits than the detailed footage it replaced.