Hacker News Reader

I found the little sketch-to-vector-variations part interesting and surprising- this is something that is solidly not done via a diffusion model unlike everything else shown. Although I note it says "We plan to build this into Firefly" implying that... this isn't something already finished.

View on HN

I imagine this would use "in the style of a line drawing" prompts under the hood to produce line-esque raster images suitable for vectorization, with the resulting vectorized images being what's shown to the user.

That was also my first instinct, but is vectorization from a line sketch really that smooth and reliable now, though? It has been some time since I've used modern tools, but last I tried any raster-to-vector on line drawings that weren't super basic, the results left a lot to be desired. Jittery, under- or over-fit, etc.

I’m intrigued as well, but especially with regards to how it would perform in the real world given that I’ve also observed that diffusion models aren’t great with vectors.

I suspect the example they used might have been cherry-picked

Reply on news.ycombinator.com