So I went away for the week-end, and my bot decided to watch porn for 48 hours.
Here are the results.
The models have been kept small for the time being, but it's based on diffusion models, the same technology that Dall-E and image-gen use,
so the current low-quality of pictures is mostly a scale issue.
I have generated ~1500 images and ~500 "videos", so server costs are minimal.
Further work will aim to allow user to generate videos from a text prompt, or from a picture.
Feel free to share the pictures on various social networks if you'd like to tickle their nude-content classifiers.
Pictures are blurry, NSFW tag is questionable due to poor image quality.
For anyone interested they were inspired by DDM but trained on 2 1080Tis over the weekend. So if you're familiar with these models you'll probably realize that the images aren't anywhere near realistic.
To the OP, maybe just keep with low image resolution (32 or maybe 64) and don't bother with 128 given your limited compute. There's no reason upscaling if you aren't training far into the 32 case.
I suspect you'll have much better results training on photos rather than videos. The individual frames of a video are often blurry, poorly composed, and uninteresting, but photos are much more carefully crafted.
"If [you] want more beach scenes, press the sand and water. If you want to see more girls in the woods, press the greenery. If you like to see girls posing on wooden chairs, press the chair. And so on. This is a solution to the problem of large-scale unsupervised image clustering," the developers explain.
15 comments
[ 2.1 ms ] story [ 42.9 ms ] threadHere are the results.
The models have been kept small for the time being, but it's based on diffusion models, the same technology that Dall-E and image-gen use, so the current low-quality of pictures is mostly a scale issue.
I have generated ~1500 images and ~500 "videos", so server costs are minimal. Further work will aim to allow user to generate videos from a text prompt, or from a picture.
Feel free to share the pictures on various social networks if you'd like to tickle their nude-content classifiers.
For anyone interested they were inspired by DDM but trained on 2 1080Tis over the weekend. So if you're familiar with these models you'll probably realize that the images aren't anywhere near realistic.
To the OP, maybe just keep with low image resolution (32 or maybe 64) and don't bother with 128 given your limited compute. There's no reason upscaling if you aren't training far into the 32 case.
Article in The Register (more than 5 years ago): https://www.theregister.com/2017/05/18/smutfinder_ai_is_now_...
"If [you] want more beach scenes, press the sand and water. If you want to see more girls in the woods, press the greenery. If you like to see girls posing on wooden chairs, press the chair. And so on. This is a solution to the problem of large-scale unsupervised image clustering," the developers explain.
Also appeared in Wired.
https://github.com/CompVis/latent-diffusion
I don't see porn, I see Hayreddin Barbarossa with the body of a roast chicken. But I guess Rule 34 exists so...maybe it is porn for someone