Hacker News Reader

Comment by daniel-thompson | Hacker News Reader

For someone only tangentially familiar with this space, how is this different than e.g. https://github.com/nerdyrodent/VQGAN-CLIP which you can also run at home? Is it the quality of the generated images?

View on HN

Stable Diffusion produces substantially higher quality images in most context, but is much more expensive to produce. The genius of VQGAN-CLIP is that it showed that you could take two pre-existing models and combine them to get text-to-image synthesis to work at all. By contrast, models like DALL-E and Stable Diffusion require extremely expensive pretraining.

There's a discussion of this in the VQGAN-CLIP paper, see in particular 6.1 "Efficiency as a Value" https://arxiv.org/abs/2204.08583

Disclaimer: I'm one of the authors of the VQGAN-CLIP paper and was tangentially involved with Stable Diffusion.

Stable diffusion generates outputs that are comparable or better than DALLE, and far beyond what VQGAN can generate.

I must say that some of my favourite images are ones that I generated in VQGAN. Don't dismiss older/smaller models. Some have a very specific quality that is perfect for some things. (Heck, I'm still fond of Aphantasia and it's weird tiling)

I personally prefer Disco Diffusion, but there's no denying stability is beyond anything else when it comes to coherency.

Yes for the end user that will be the main difference. From the curation of the training data to the model itself a number of things have been put together that make the generations substantially more aesthetically pleasing imho

Reply on news.ycombinator.com