stefatorus

On PNGs, the benefit is smaller and mainly perceptual. dE00 actually decreases with the SR model but perceptual quality increases especially if the original raster was very low resolution.

On JPEGs or other lossy compression inputs, both models help significantly. They've been trained with slightly different objectives. The SR one uses standard SR loss (with extra weighting near borders), the denoiser is trained to enforce the axioms that the vectorizer uses so it's the one that's recommended for production.

I haven't tried running both at the same time, but it might be worth experimenting. Both are open weight so you can check yourself if curious.

https://huggingface.co/Logolabs/inkvec-denoiser-001 https://huggingface.co/Logolabs/inkvec-sr-001

Novel engineering contribution implementing several papers to improve the quality of raster-to-vector conversion.

From our internal testing, it's the open-source SoTA, surpassed only by vectorizer.ai.

It is built in Rust, has a wasm based version you can play with, and 2 models (one for SR, one for denoising) to help it work better with in-the-wild images.

https://huggingface.co/spaces/Logolabs/inkvec

The trick to getting a model to perform on something is to have it as a training data subset.

OpenAI might have thought Chess is good to optimize for but it wasn't seen as useful so they dropped it.

This is what people refer to as "lobotomy", ai models are wasting compute on knowing how loud the cicadas are and how wide the green cockroach is when mating.

Good models are about the training data you push in em