Thanks! No, I use banana.dev because I currently have too little traffic to run the GPUs 24/7. That's why the initial generation is currently a little slow, but hopefully that will be improved with some new changes they're implementing soon.
Automatic1111's fork is pretty good in the no-messing-around sense – on my Windows box (which has an old but decent gaming GPU), all it took was running `launch.bat`, and it set up a venv, downloaded dependencies and wired everything up, and opened the Gradio web UI.
This is why things like reddit / hn exist; a raw list of links gives you no idea what’s interesting and what’s just irrelevant.
I think the tldr is, in the last month:
- 1.5 model came out; its ok. Incremental improvement, not really significant.
- new VAE came out; this actually tangibly improves fine details, like feet and hands.
The rest is random crap around supporting tooling or vague hand wavy research stuff.
Don’t get me wrong, lots of stuff happening, that’s great, but a lot of it is basically nothing really worth paying attention to unless you’re specifically invested in the topic.
The 1.5 inpainting model is great! Significant improvement over how to do inpainting before. Highly recommended. Integrated into my mobile app last night and enjoying it so far!
any idea if you can merge a dreambooth checkpoint with the inpainting model to get the best of both? I mostly use inpainting to fix custom trained faces.
Yeah, that would be tricky to merge existing ones together. But you should be able to fine-tune the inpainting model the same way as you tune with dreambooth?
For faces, I haven't looked deep, but seems CodeFormer's training cost is minimal, should be able to fine-tune that model, probably better?
23 comments
[ 2.8 ms ] story [ 59.2 ms ] threadhttps://www.phantasmagoria.me
Are you running the GPU instances and model yourself, or are you outsourcing image generation to an API?
So I used this and it's as simple as extract the download and run the curl | bash script inside it.
Is it safe for enterprise use? Nay.
Is it more convenient than manually futzing with venvs? Yay.
https://github.com/cmdr2/stable-diffusion-ui
I made a service [1] to use SD in Notion. Is it something you could share as well?
[1] https://slashdreamer.com/
I think the tldr is, in the last month:
- 1.5 model came out; its ok. Incremental improvement, not really significant.
- new VAE came out; this actually tangibly improves fine details, like feet and hands.
The rest is random crap around supporting tooling or vague hand wavy research stuff.
Don’t get me wrong, lots of stuff happening, that’s great, but a lot of it is basically nothing really worth paying attention to unless you’re specifically invested in the topic.
For faces, I haven't looked deep, but seems CodeFormer's training cost is minimal, should be able to fine-tune that model, probably better?
[1] https://openart.ai/promptbook