Chess Solver from Image: How AI Reads Your Board and Finds the Best Move

Typing a position into a chess engine square by square takes longer than most positions are worth. A chess solver from image removes that step: you hand it a screenshot, it works out what is on the board, and it hands back the strongest move. The technology is now reliable enough that the interesting question is no longer whether it works — it is which approach you want, because the tools in this category differ far more than their store listings suggest.

This guide covers how image recognition turns a picture into a playable position, what actually makes a scan fail, how the cloud and on-device approaches differ in ways that matter to you, and where the whole thing sits relative to the rules.

How Chess Image Recognition Works

Turning a picture into a position is a pipeline, and each stage fails in its own way:

1. Board detection. The software has to find the eight-by-eight grid inside an image that also contains a clock, a move list, an opponent's avatar and possibly an advert. It does this by looking for the checkerboard pattern — a regular lattice of alternating light and dark squares.
2. Grid mapping. Once located, the board is divided into 64 cells, each given a coordinate from a1 to h8.
3. Piece recognition. Each cell is classified: empty, or one of twelve piece types. A neural network does this, and it can only recognise piece designs it has been trained on.
4. Orientation. Is White at the bottom, or are you playing Black and looking at a flipped board? Getting this wrong produces a legal-looking position that is the mirror image of reality — a worse failure than an obvious error, because nothing looks broken.
5. FEN conversion. The result is written as FEN (Forsyth-Edwards Notation), the standard one-line text format for a chess position.
6. Engine analysis. The FEN goes to a chess engine, which returns the strongest move and an evaluation.

On a modern phone the whole sequence takes a couple of seconds. Worth knowing about the last step: these tools analyse the position in front of them. You get one move, not a script. Play it, and the new position is read again and the next move arrives — which is how you walk a position through to checkmate.

{{APP_DOWNLOAD_SECTION|Solve any chess position from a screenshot in seconds. Our AI reads your board and finds the strongest move instantly!}}

What Actually Makes a Scan Fail

Very few failed scans are interesting. Four things cause the overwhelming majority, and all four are avoidable.

The board is cropped

A board missing a rank cannot be reconstructed — the information is simply not in the image. All 64 squares need to be in frame. Cropping tight to the board edge is fine; cropping into it is fatal.

Something is sitting on top of the squares

A move dialog, a "your turn" notification, an advert banner, a piece mid-drag, a cursor. Anything overlapping the board hides squares, and a hidden square is a guess. Take the shot when the board is quiet.

It is a photograph of a screen rather than a screenshot

Photographing a monitor adds glare, moiré, keystone distortion and a colour cast, and every one of those makes recognition harder. If the position is on a device you control, use the screenshot function. This is the easiest accuracy gain available and costs nothing.

The board theme has low contrast

This is the one nobody thinks of, and it matters more than the other three combined. Board detection depends on the checkerboard pattern being visible, and some themes make it nearly invisible to software: translucent boards tinted by whatever sits behind them, high-saturation neon palettes where the two square colours have almost identical brightness, and heavy wood or marble grain where the texture inside a single square varies more than the difference between two adjacent squares.

There is a genuinely useful test for this. Squint at the board with your eyes half closed. Squinting destroys fine detail and leaves large-scale brightness differences, which is roughly what the detection stage works with. If the grid is still obvious, software will find it. If the board flattens into a rectangle, expect trouble — and switching to a plain high-contrast theme for a moment turns an unreliable scan into a reliable one.

The same logic applies to piece sets. Recognition is trained against specific designs, so a common, clean set scans better than an ornate or unusual one. No scanner reads every board and piece combination in existence, and any tool claiming otherwise is overselling.

3D Boards Do Not Work

No mainstream scanner reads 3D board views, and this is a limit of the problem rather than of the software. In a 3D view pieces occlude each other — a rook on the back rank can be entirely hidden behind a queen in front of it, and no amount of model training recovers information that is not in the image. The grid is also foreshortened, so distant squares are a fraction of the size of near ones and the pieces on them are drawn smaller.

If you play in a 3D view, switching to 2D for the moment you take the screenshot is the whole fix. The position is identical; only the rendering changes.

Cloud or On-Device: the Choice That Actually Matters

Every tool in this category makes one architectural decision that shapes everything else about it. Either the picture leaves your phone or it does not.

Cloud tools upload your screenshot, run recognition and analysis on a server, and send the result back. That buys enormous compute — server-grade engines searching far deeper than a phone can — which is why the most powerful analysis platforms work this way. The costs are that it needs a connection, it usually needs an account, it typically runs on a subscription because servers cost money every month, and your positions pass through somebody else's machine.

On-device tools run everything locally. Search depth is bounded by your phone rather than a data centre, but nothing is uploaded, no account is required, it works with no signal at all, and the pricing does not have to be a subscription because there is no recurring server cost to cover.

For the way most people use these tools, the depth difference is theoretical. Engine strength has sharply diminishing returns: the first ten plies of search catch the hanging pieces and short tactics that decide the overwhelming majority of amateur games, and past roughly depth twenty the evaluation keeps drifting while the recommended move stops changing. A phone searching two seconds and a server searching two minutes will usually hand you the same move. They will disagree about whether it is worth +1.2 or +1.4, which changes nothing about what you should play.

Where server depth genuinely earns its keep is sharp, heavily tactical positions and long forcing sequences. In quiet positions the extra search is buying decimal places.

How the Main Options Compare

The established tools split cleanly along that line.

• Chessvision.ai Chess Scanner — the closest direct competitor. Broad ecosystem including a browser extension and ebook reading, server-side processing, membership billed monthly or annually with no one-time option.
• Chessify — aimed up-market at serious and titled players. Scans boards, book diagrams and screens, then runs deep multi-engine cloud analysis alongside databases and video search. Metered cloud compute on top of a subscription. Powerful, and considerably more tool than a "what do I play here" question needs.
• Chess Move (Stockfish Engine) — camera-first rather than screenshot-first, cloud engine, subscription after a short free trial.
• Chess AI Solver — on-device recognition and analysis, no account, works offline, free tier with a one-time unlock rather than a recurring charge.

Which is right depends entirely on what you are doing. If you want multi-engine analysis across a database of your tournament games, a cloud platform is the correct answer and no phone will match it. If you want to know what to play in the position currently on your screen, without an account and without a monthly charge, the on-device tools are built for exactly that.

{{APP_DOWNLOAD_SECTION|No account, no connection, nothing uploaded. Scan a position and get the best move right on your phone.}}

Screenshot Analysis vs. Entering a FEN

Every scanner also accepts a FEN string directly, and there are times when typing one is the better option.

Screenshots win whenever the position is already on a screen. They are faster, they carry the whole position including castling rights and side to move, and there is nothing to mistype.

FEN wins in three cases: when the position comes from a source no scanner reads cleanly, such as a 3D view; when you already have the FEN from another tool and pasting it is instant; and when you are constructing a hypothetical position that does not exist on any screen. FEN is also worth learning for its own sake — it is the interchange format the whole chess software world runs on, so reading one makes every other tool easier to move between.

Is Using a Chess Solver Cheating?

It depends entirely on when you use it, and the distinction is not subtle.

Analysing your own finished games, working through puzzles, studying an opening line, checking whether the move you played was the best available — this is how every strong player has studied for the last thirty years. Engines are the standard tool of chess improvement, and nobody considers this cheating.

Using engine assistance during a live rated game against another person is against the rules of Chess.com, Lichess and essentially every other platform and federation. Accounts are closed for it, permanently, and appeals rarely succeed.

Fair-play enforcement works by reviewing the games themselves afterwards, and it runs whether or not a screenshot was taken. There is no careful way to use engine assistance during a live game. If you want to keep your account, keep the engine for positions that are no longer live.

Frequently Asked Questions

How does a chess solver read an image?

It finds the board by looking for the checkerboard pattern, divides it into 64 squares, classifies what sits on each one with a neural network, works out which way round the board is, and writes the result as a FEN string that a chess engine can analyse.

Can I upload a chess screenshot to get the best move?

Yes — that is the core function of every tool in this category. Where they differ is whether the upload leaves your device. On-device tools read the screenshot locally and transmit nothing; cloud tools send the image to a server for processing.

Is there a chess engine that works from a screenshot?

Yes. The screenshot is converted to a FEN position and that position is handed to a conventional chess engine. The engine is doing ordinary chess analysis — the image recognition is a front end that saves you typing the position in.

Can I use my camera instead of a screenshot?

Some tools are camera-first. Printed diagrams in books and magazines are the most reliable camera case, because the page is flat and matte and the diagram is drawn in 2D. Photographing a screen works far less reliably than screenshotting it. Photographing a physical board with three-dimensional pieces is a substantially harder problem — the pieces hide one another and the perspective distorts the grid — and most scanners do not attempt it.

Can it solve chess puzzles from images?

Yes, and it is one of the better uses. Puzzle positions are usually clean, well framed and 2D, which is exactly the case recognition handles most reliably.

How accurate is AI chess image recognition?

On a common board theme with a common piece set, very. The everyday failures are cropping, obstruction and low-contrast themes rather than genuine misclassification. Accuracy drops on unusual or heavily textured skins and on piece sets a model has not been trained against, and no scanner covers every combination that exists.

Does the position have to come from Chess.com or Lichess?

Those two are what these tools are principally built and tested around, so they are the best-supported case. Screenshots from other apps and sites often work, and book diagrams frequently do, but coverage outside the two major platforms varies by tool and is worth testing before you rely on it.

Do I need an internet connection?

For cloud tools, always — they cannot function without one. For on-device tools, no: recognition and analysis both run locally, so they work on a plane or with no signal.

What does the evaluation number mean?

It is measured in pawns but it is not counting them. +1.4 means the engine judges White to be about a pawn and a half better once king safety, piece activity and pawn structure are weighed together. Under ±0.3 is level; beyond ±3 is winning given accurate play. When the display switches to something like M5, the engine has stopped estimating and has proved a forced mate in five.

Why does the engine change its mind about the best move?

Because it is searching deeper. An evaluation is always the value of the position at the end of the line calculated so far, not a verdict on the position in front of you. Search two moves further and a refutation can appear that reverses the judgement. Early evaluations are provisional by nature.

What is the best chess screenshot solver app?

There is no single answer, because these tools optimise for different things. For deep multi-engine analysis across a game database, a cloud platform is built for that. For a position solved on your phone with no account, no connection and no subscription, an on-device tool is built for that instead. Decide which of those describes you before comparing feature lists.

Are these tools free?

Most offer something free and charge for volume or for advanced analysis. The meaningful difference is the shape of the charge: cloud tools generally have to bill on subscriptions because their server costs recur every month, while on-device tools can offer one-time purchases because they do not.

The Short Version

Reading a chess position from an image is a solved problem in the ordinary case and a contrast problem in the difficult ones. Give it the whole board, unobstructed, in 2D, on a theme where the grid is clearly visible, and it will work. Then choose your tool on the axis that genuinely differs between them — whether your positions leave your device, and whether you would rather pay every month or once.