stylegan VS DeOldify

Umm, StyleGAN was the first decent image generation model, and it was producing great images from random seeds 5 years ago. Now, that's with the obvious caveat that each model was trained to produce one specific type of image and it helped immensely if the training images were all aligned the same. Diffusion models are certainly the trendy current architecture for image generation, but AFAIK there's no fundamental theoretical limitation to the output quality of any architecture except the general rule that more parameters is better.

PDF link Landing page

Code: https://github.com/NVlabs/stylegan

https://arxiv.org/abs/1812.04948 (A Style-Based Generator Architecture for Generative Adversarial Networks)

Check out the PDF here I've been calling this "the first spiritual guidebook written by an AI". The text was written almost entirely with GPT-3, minus the couple of paragraphs I wrote as prompt. The art was made with various tools - for instance, the cover was made in part with Aphantasia (CLIP+FFT), the title page is an old public domain photo colorized with DeOldify, the faces made with StyleGAN, both standalone and via Artbreeder. Let me know what you think - I have a couple bigger projects building off of the techniques I used, and I would appreciate any feedback!

stylegan this is the tech behind all those "this ___ does not exist" sites. it's actually surprisingly easy to use if you know how to use the terminal (and preferably a little bit of python).

Clone stylegan

The old method is simple. All you need to do is running DeOldify with a colab notebook. Then you can pick any neural networks you want for upscale and frame interpolation. A lot of people use dain-app, but I find rife faster and easier to use.

I think the novel and interesting tech is still happening, its just that without the colorful ads for it on TV, and without the software being packaged up and sold with pretty box art that you can physically hold, it doesn't feel as much like a capital-E Experience. It's probably the Internet's fault that we don't do things like that anymore, but the upside is that we now have access to so many ideas and applications from all over, even ones that aren't commercially viable.

Some that look exciting to me are: an AI that lets you animate still photos realistically [1], a simple website that guides you to discover new parks, eateries, and other places near you [2], an AI that colorizes old black-and-white photos/video [3], a Street View style map of the game world from "The Legend of Zelda: Breath of the Wild", with some 1st person 360 degree photos [4], and a tiny game engine that lets you distribute your whole game physically via printed QR codes [5].

If marketing and graphic design people ever felt like getting together to do some 'side projects', I vote that they should make print ads for apps/websites that they like :)

[1] https://github.com/AliaksandrSiarohin/first-order-model

[2] https://randomlocation.xyz (https://randomlocation.xyz/help.txt for customization)

[3] https://github.com/jantic/DeOldify

[4] https://nassimsoftware.github.io/zeldabotwstreetview/

[5] https://github.com/kesiev/rewtro

Thank you. Photo-restore + Deoldify + Animegan + Rife + 4K upscale

DeOldify

I don't know if that's a great example of the benefits of domain experts for ML. She specifically cherrypicked images which are themselves extremely unusual for old photographs (and you've probably seen them shared on social media before precisely because they are such colorful pioneering color photographs, with a level of quality & detail that wouldn't become common in color photography for a long time afterwards). And the point about minimizing average error tending to make for conservative image choices rather than sampling from the distribution is an old one which is familiar to anyone with the most passing interest in NN colorization - if you look up the (now 3-4-year-old) library used, Antic's DeOldify, his discussion is pretty much all about how to fight the averaging using GAN techniques which encourage the colorization model to at least pick a mode and get nice bright colors that way while minimizing GAN use as much as possible (because GANs are miserable to train). You don't need a Twitter 🧵 lecturing you on this deep insight, everyone is already well-aware of that the first time they train a decolorizer and go "why is it so brown".