Text-to-image models can create realistic pictures of clocks but cannot reliably set the time they show, often defaulting to a common pose like 10:10. The same problem occurs when asked to show a specific fraction of a round object, like a pizza slice: the model returns a whole pizza or a standard cut instead of the requested amount.
The researchers identified that both tasks involve controlling angles: the positions of clock hands and the boundaries of a slice. They designed a method that encodes the target angle as a pair of numbers (sine and cosine) and adds this information as special tokens in the text prompt, without needing a diagram or bounding box at the time of generating the image.
On the clock task, the method achieved accurate hand placement between 63% and 78% of the time, far above the less-than-1% accuracy of standard models. For circular fractions, accuracy rose from roughly 20% to 89.5%, and the approach generalized to new object shapes and target values it had not seen during training.