Hacker News new | ask | show | jobs
by rookhack 1352 days ago
We haven't tried any yet, but would love to. Stable Diffusion for images would be super cool. Which would you recommend?
1 comments

I was curious about how GPT-NeoX would match up to GPT-3. It seems like the most capable of the freely available models (at least based on their specs). Of course more parameters != quality of their dataset, so I can't really conclude much in this scenario.

As far as I know, https://textsynth.com is the cheapest host and NeoX-20B model is cheaper than OpenAI's Curie. They have a Playground which allows you to experiment without an account.

Re Stable Diffusion:

Based on the results I've seen in https://news.ycombinator.com/item?id=33038117 and other services, if you just want some eye candy I think Stable Diffusion would work fine (and is much better looking than anything DALL-E could generate).

However I wasn't impressed by SD's ability to understand the meaning of my sentence, it seems to just create an image by mashing words together (and sometimes even ignores key words/meanings). According to commenters, DALL-E is much better at "understanding" sentences so if that matters for this site; then I guess you should avoid SD.

awesome - appreciate that! Will check out textsynth.

We would LOVE a way to generate logos effectively, maybe Stable Diffusion is the answer there. Right now DALL-E really struggles with that.

Logos are really hard though.

Check out the free, open source https://stablehorde.net/ for stable diffusion image generation via a distributed cluster (with REST API).
I disagree with this, I don't think it's a good look to use a free service with limited resources for something commercial; it's just bound to cause conflicts.

But this service (and their sister project for text generation) is incredibly cool and interesting, surprised I haven't heard about it sooner.