Each point of a vector drawing is stored as one token
A logo file is a small program for a pen. Go here. Curve toward there. Fill the shape. The numbers are the drawing. That is why the edge stays sharp when the logo grows, and why a person can still grab a point and move it. Those numbers are also what a language model guesses worst. The model continues text. A coordinate is text, and nothing in a digit keeps it on the page. The authors' name for the miss is coordinate hallucination.
Yiying Yang, Wei Cheng and their colleagues, at Fudan University and at StepFun, try to stop the miss. The paper is OmniSVG: A Unified Scalable Vector Graphics Generation Model, from the 2025 Conference on Neural Information Processing Systems. They take the drawing off the page of text. Every picture in the training set is pressed into a box of 200 by 200. A tool called picosvg throws out the parts that keep a design file in order: groups, the transforms that move those groups, and the spare attributes. Five pen moves remain. Move. Line. Curve. Arc. Close. A sixth move carries the fill, so the colour is a choice of its own. The paths are then laid in one strip, from start to end.
A point on that box is no longer two numbers to spell. It is one square. The name of the square is x × w + y, and the width is 200. A sheet that size has 40,000 squares. Their figure marks the drawing's vocabulary at 40,000, beside the 150,000-word vocabulary of the language model they start from, Qwen2.5-VL. That model already continues a sentence by picking the next word. They give it a second alphabet, made of squares and of those few pen moves, and the new alphabet is trained from scratch on top of the old one. The next guess is a place the pen can actually land.
They test the alphabet by taking it away. With coordinates and colours both written as ordinary text, a sample ran to 18.5 thousand tokens, and the pictures sat furthest from the reference set. Packing the colour and leaving the coordinates as text cut that to 10.2 thousand. Packing the coordinates and leaving the colour cut it to 6.3 thousand. Packing both left 4.8 thousand, and the pictures sat closest. The other scores in the same table move in that order. The coordinates were the long part, and they were the part free to wander. A square shortened the file and drew the pictures toward the reference set. One change did both.
The rows belong to the model they call OmniSVG, with 3.7 billion parameters, built on Qwen. An arc is the leftover. It still needs two radii, an angle and two flags, and those are not squares. The formula never claimed them. What comes out is the strip written down again as paths. The pen moves, the shape closes, a colour fills it. A group that went into the cleaner does not come out, because the model was never given a word for a group. A dense figure still costs tens of thousands of these choices, and the authors call the wait considerable. The squares got the numbers under control. A detailed drawing is still one choice for every place the pen stops. Show the model a photograph and it fails. They write that it follows only pictures that already look like vector art. It learned to continue a path. A camera does not leave one.