[yklcs]

I’ve tried out some work on generating vector fonts too, in the format of Bezier curves and a seq2seq model. The problem was that fonts outputted by ML models were imprecise. Lines were not perfectly parallel, corners were at 89°, and curves were kinked. It’s not too difficult to get fonts that look good enough, but the imperfections are glaring as fonts are normally perfectly precise. These imperfections are evident in OP’s output too, and in my opinion make these types of models unusable for actual typesetting.

A 1% error in a raster output would be pixel colors being slightly off, but a 89° corner in a vector image is immediately noticeable, making this a hard problem to solve. I haven’t looked into this problem too much since, but I’m interested to hear about possible solutions and reading material.

[waterheater]

Without changing the fundamental learning process, one could conceivably introduce a "post-production" step, where you tighten up the output according to a set of pre-defined rules (e.g., if an angle is 89 degrees, adjust the angle to 90).

Of course, changing the learning process would be best. One idea which comes to mind is finding a way to embed relationships into the ML training system itself (e.g., output no angles other than 90 degrees or some predefined set). Such an approach is a type of contraint-based ML, where the ML agent identifies a solution given certain constraints on the output. In my experience, the right approach to accomplish this goal is using factor graphs.