I investigated a gradient-descent-based concept adjacent to this back in like 2023 [0] inspired by the original 3DGS paper, I was pretty happy with the results I got with back-to-front alpha-blending and polar-coordinate gaussians for variable-curvature brush strokes. One of these days I should add another example gif to the repo besides just the Mona Lisa one that's currently there, but that's the only one I've got at the moment unfortunately.
As someone with an arts degree (not in painting though, but still) it feels really really weird to see how that animation builds up the image. It feels less random than the old "simulated annealing with vector shapes" techniques, ending up somewhere that feels both similar and extremely far removed from how a human would do it at the same.
It's like… the way a trained human would paint feels kind of like how MSB radix sort makes an array look more ordered, going from crude to refined. The way your algorithm builds up an i|age feels like I'm looking at a radix sort that alternates between a stable MSB and LSB radix sort, which is weird as heck. It's fascinating, really.
The results look way better than I was expecting. I don't know anything about Gaussians, but most of the "painting" effects I tried in image editors didn't produce satisfying results. This one looks great.
The effect looks decent in the foreground of some pictures, particularly the wolf and the cat. It really exaggerates the depth of field in the background, though, and those areas feel like someone ran a posterise after a blur, rather than brush strokes. Might just be a side effect of the author's predilection for bokeh, I guess. The Tokyo shot looks good throughout.
Intriguing and nice writeup. This however tries to encode all the hidden painter's knowledge. Would it be easier to generate realistic images from paintings and use these pairs to finetune an image generation model?
> In the days of generative image models, it's still fun to see what can be achieved using traditional image transformation techniques.
> But the really fun part of the project was in combining a number of techniques, such as Perlin noise, wavelets, and edge detection, that I played around with previously in isolation. All these different tricks came together for this project, making it possible to build something greater than the sum of its parts. I find these are the most rewarding types of experiments where you can build on things you've previously learned and combine them in novel ways to make something new and unexpected. I hope you enjoyed the journey as much as I did working on the project.
You can already use genAI to do similar things. The whole point, however, was to see what I could do using traditional techniques. The other benefit is that it's something you can run locally without needing a lot of resources, and have full control over the shape of the output. You can also combine this with generative models, and have a model generate a picture, then have this repaint it. Ultimately the point was to have some fun and play around with combining these techniques.
I used my own photos for the input, I find images with the subject isolated to be more visually pleasing, hence most of my photos are in this style. But there wouldn't be any difference to the algorithm if the image was full of detail as you can see with my photo of Hong Kong at the end.
I investigated a gradient-descent-based concept adjacent to this back in like 2023 [0] inspired by the original 3DGS paper, I was pretty happy with the results I got with back-to-front alpha-blending and polar-coordinate gaussians for variable-curvature brush strokes. One of these days I should add another example gif to the repo besides just the Mona Lisa one that's currently there, but that's the only one I've got at the moment unfortunately.
[0] https://github.com/mkaic/painterbot
As someone with an arts degree (not in painting though, but still) it feels really really weird to see how that animation builds up the image. It feels less random than the old "simulated annealing with vector shapes" techniques, ending up somewhere that feels both similar and extremely far removed from how a human would do it at the same.
It's like… the way a trained human would paint feels kind of like how MSB radix sort makes an array look more ordered, going from crude to refined. The way your algorithm builds up an i|age feels like I'm looking at a radix sort that alternates between a stable MSB and LSB radix sort, which is weird as heck. It's fascinating, really.
Can you make the brush strokes look like brush strokes?
Neat, watching it paint as an animation is pretty fun.
The results look way better than I was expecting. I don't know anything about Gaussians, but most of the "painting" effects I tried in image editors didn't produce satisfying results. This one looks great.
The effect looks decent in the foreground of some pictures, particularly the wolf and the cat. It really exaggerates the depth of field in the background, though, and those areas feel like someone ran a posterise after a blur, rather than brush strokes. Might just be a side effect of the author's predilection for bokeh, I guess. The Tokyo shot looks good throughout.
It was more of a stylistic choice on my part. I could add more layers to flesh out background details. And it's Hong Kong actually. :)
Intriguing and nice writeup. This however tries to encode all the hidden painter's knowledge. Would it be easier to generate realistic images from paintings and use these pairs to finetune an image generation model?
From the article
> In the days of generative image models, it's still fun to see what can be achieved using traditional image transformation techniques.
> But the really fun part of the project was in combining a number of techniques, such as Perlin noise, wavelets, and edge detection, that I played around with previously in isolation. All these different tricks came together for this project, making it possible to build something greater than the sum of its parts. I find these are the most rewarding types of experiments where you can build on things you've previously learned and combine them in novel ways to make something new and unexpected. I hope you enjoyed the journey as much as I did working on the project.
You can already use genAI to do similar things. The whole point, however, was to see what I could do using traditional techniques. The other benefit is that it's something you can run locally without needing a lot of resources, and have full control over the shape of the output. You can also combine this with generative models, and have a model generate a picture, then have this repaint it. Ultimately the point was to have some fun and play around with combining these techniques.
I’d suggest for the examples to pick images that are not mostly bokeh.
I used my own photos for the input, I find images with the subject isolated to be more visually pleasing, hence most of my photos are in this style. But there wouldn't be any difference to the algorithm if the image was full of detail as you can see with my photo of Hong Kong at the end.
Super impressive results, and an awesome write up. Thanks :)
Glad you liked it!
Very cool
Thanks, it was a really fun thread to pull on. :)