Category: 05-ArtBreeder

It’s alright. There are still some deconvolution artifacts that give it away immediately (at least with the more photorealistic genres I’ve played with.) I love the freedom and flexibility it provides with various keywords and parameters, but the actual effect of these parameters in the images is rather disappointing. Resolution is not bad for generative models.

Koke_Cacao-ArtBreeder

(05-ArtBreeder)
ArtBreeder
ProsePainter

The art breeder model is very cool. It magically preserves the lower-level pattern of oil paintings with relatively
high resolution. This quality is often, at least to me, hard to achieve and requires a lot of hyperparameter tuning
and GPU power. The fact that it can run on the browser this fast is amazing (is it running on a browser though? Or
does it use backend GPU with querying?). Also, it is very rare to have a user interface that allows both breeding
and gene-editing. I wonder if “breeding” means parameter breeding rather than pixel-wise breeding.

Artbreeder

Words used: Butcher shop, nipple, hen, velvet, butterfly, chaos

These are the images I ended up generating with ArtBreeder. I gravitated towards creating somewhat bleak landscapes, and I think a lot of the landscapes ArtBreeder generates are really interesting. I could see myself using the tool to help generate the base aesthetics of some alien/sci-fi/alternate dimension landscape. Also, I found it super surprising that this was made by a BCSA almunus. Although I hadn’t used ArtBreeder before,  I have definitely seen it talked about all  over the internet. I’m surprised that CMU doesn’t advertise this fact front and center on the BSCA info page.

These are two images  I generate, using the shoe and cat modes respectively. I thought this was pretty cool, but at the same time pretty frustrating, because with the cat one I gravitated to making the cat have really complex, unnatural forms, but Pix2Pix kept having trouble interpreting the image and it would keep coming out as an incomprehensible, blurry mess. I do realize why this is the case, however.

It’s actually really hard to imagine what the outcome will be just based on the sliders. The resulted images aren’t really what I was  expecting, but they do have somewhat distinguishable features from the genes.

Really cool and fun to play with. I can see the texture of the “ball” is getting furrier and greener as I raise the gene proportion of “tennis ball”.

This was pretty fun but I thought once the amount of genes got past 4 or 5, the outcome became significantly less controllable. In the second image I had pomeranian as one of the genes but I kept getting images of babies.

The interface takes a little getting used to (I didn’t realize the “Parents” I was initially selecting were actually generated images themselves) but is addicting once you get going. Fun to think of inputs that visually kind of work together (or not) but come from very different places. For some reason I wanted to start from real rather than generated images, and kept it to 2-3 inputs so their sources were still somewhat recognizable.