A tiny tool for precise pixel edits
Teach an edit.
Catch the guess.
One example can mean three different things. Teach a reusable sprite edit, see every explanation that fits, and settle the difference before it touches a batch.
Try the live exampleNo account · No model download · Original sample art
THE WORKBENCH
From example to evidence.
No autosave. Export evidence JSON and check the file before leaving. A browser leave warning is best effort and may not appear, especially on mobile or after a crash. A download request does not confirm a saved file.
Show the edit
8 × 8 pixelsPaint a before and after. Arrow keys move through a canvas; Space paints.
Before
After
Inspect the possibilities
Ask a better question
Preview the batch
A result appears only when every surviving rule gives the same output. Hatched pixels mark disagreement, not a finished edit.
Restore a saved session
Open an exported evidence JSON file as text and paste its contents below. Version 1 and 2 files are supported. Only the 8 × 8 examples and three batch inputs are restored; version 2 also supports sessions with no learned examples. Predictions are recomputed here. No upload or network request.
This replaces your current saved examples, batch inputs, and any unsaved canvas edits. To keep your current work, cancel first, save or learn any pending edits, then export your evidence.
SMALL AI. VISIBLE REASONING.
No magic.
No silent guesses.
Pixel Witness uses symbolic AI: it searches 176 component-aware programs, keeps the ones that explain your examples, and chooses a question whose possible answers split those programs apart.
That is inductive program synthesis and active learning. It does not use an LLM, a pretrained network, or a remote service.
What the learner can express
One component selector followed by one action. Select by color, size, shape proportions, or canvas edge. Recolor, erase, outline, or move one pixel. Components are same-color pixels joined horizontally or vertically; touching different colors are separate.
What it cannot promise
Only 8 × 8 grids with three ink colors are supported. There is no semantic object recognition, multistep recipe, tolerance for mislabeled pixels, or proof that your intent lies in this language. Two different intentions may agree on every example here. A unique recipe is not certainty about your intent.
Precise editing behavior
Size selectors include every tied component. Outlines use four neighboring pixels, clip at edges, and never overwrite existing ink. Moving is atomic: the whole move stays unchanged if any selected pixel would leave the canvas or collide with unselected art.
Prior art and prototype status
Learning image edits and asking targeted questions are established ideas. This original prototype adapts them to a small, inspectable pixel-art workflow. See ImageEye, SmartLabel, and Pattern Analogies. Synthetic tests do not establish artist productivity or accuracy on real sprite collections.