4 ms·

GitHub Repo: https://github.com/Stability-AI/stablediffusion https://github.com/Stability-AI/stablediffusion HuggingFace Space (currently overloaded unsurprisi

GitHub Repo: https://github.com/Stability-AI/stablediffusion https://github.com/Stability-AI/stablediffusion HuggingFace Space (currently overloaded unsurprisingly): https://huggingface.co/spaces/stabilityai/stable-diffusion https://huggingface.co/spaces/stabilityai/stable-diffusion Doing a 2.0 release on a (US) 2-day holiday weekend is an interesting move. It seems a tad more difficult to set up the model than the previous version.

gst 4y ago
> Doing a 2.0 release on a (US) 2-day holiday weekend is an interesting move. Their HQ seems to be located in London.
- minimaxir 4y ago
  True, but it's going to dampen the launch a bit.
  - jehb 4y ago
    Is it? For me, I'm in tech but nowhere near anything for which playing with a new SD release would be relevant to my day job. Having a couple of extra days off to play with a new tech toy probably means I'll use it more.
  - smcleod 4y ago
    I don't see why? The world is a whole lot bigger than the USA!
Operyl 4y ago
Seems like a potentially good time to launch it, lots of young people with free time.
- Gigachad 4y ago
  Less likely journalist whingers will pick it up too.
- kennyloginz 4y ago
  Definitely something to talk about or share with the fam.
Existenceblinks 4y ago
I wonder why these AI repos' documentation are so bad compared to what we are used to in general. Where is intro/get started/example(commands)/config(docs) etc.
- blihp 4y ago
  I believe it's a combination of the fact that most of these models are basically 'research dumps' primarily targeting other researchers and given this they are assuming a level of familiarity with related tools/libraries. So it's up to interested people in the community to take it the last mile/block/whatever to make it easy to use, address specific use cases etc. for use by a less academic/technical audience.
  - Existenceblinks 4y ago
    Yeah, generally speaking, it's not optimized for open collaboration (not even close .. pun intended).
IanCal 4y ago
I think it's looking fairly similar, the first one was a bit tricky too. Later improvements by the community made it clearer. The docs aren't good though, it tells you to download two things when actually I think you only need one. If you do need two then it doesn't tell you at all where to put the second. You really need xformers if you're doing it at home, I've got a 3090 and it blew through the ram without it. However, the instructions didn't work for me for compiling and there's an incompatibility if you try and install from conda. You can have it work but you need to upgrade python from 3.8.5 to 3.9 in the yaml file first, then you can install it (xformers needs 3.9+, and something else in SD breaks on 3.10+ so 3.9 works). This needs the classic "sit next to a new person installing it by following the docs and see what problems they hit, fix the docs and start from scratch again" process. Looks good, though so far the images I've made don't look as nice as with 1.4, but I guess that's largely down to finding the right tweaks for the model and right magic wording for the prompts.

gst 4y ago

> Doing a 2.0 release on a (US) 2-day holiday weekend is an interesting move. Their HQ seems to be located in London.

minimaxir 4y ago
True, but it's going to dampen the launch a bit.
- jehb 4y ago
  Is it? For me, I'm in tech but nowhere near anything for which playing with a new SD release would be relevant to my day job. Having a couple of extra days off to play with a new tech toy probably means I'll use it more.
- smcleod 4y ago
  I don't see why? The world is a whole lot bigger than the USA!

Operyl 4y ago

Seems like a potentially good time to launch it, lots of young people with free time.

Gigachad 4y ago
Less likely journalist whingers will pick it up too.
kennyloginz 4y ago
Definitely something to talk about or share with the fam.

Existenceblinks 4y ago

I wonder why these AI repos' documentation are so bad compared to what we are used to in general. Where is intro/get started/example(commands)/config(docs) etc.

blihp 4y ago
I believe it's a combination of the fact that most of these models are basically 'research dumps' primarily targeting other researchers and given this they are assuming a level of familiarity with related tools/libraries. So it's up to interested people in the community to take it the last mile/block/whatever to make it easy to use, address specific use cases etc. for use by a less academic/technical audience.
- Existenceblinks 4y ago
  Yeah, generally speaking, it's not optimized for open collaboration (not even close .. pun intended).

IanCal 4y ago

I think it's looking fairly similar, the first one was a bit tricky too. Later improvements by the community made it clearer. The docs aren't good though, it tells you to download two things when actually I think you only need one. If you do need two then it doesn't tell you at all where to put the second. You really need xformers if you're doing it at home, I've got a 3090 and it blew through the ram without it. However, the instructions didn't work for me for compiling and there's an incompatibility if you try and install from conda. You can have it work but you need to upgrade python from 3.8.5 to 3.9 in the yaml file first, then you can install it (xformers needs 3.9+, and something else in SD breaks on 3.10+ so 3.9 works). This needs the classic "sit next to a new person installing it by following the docs and see what problems they hit, fix the docs and start from scratch again" process. Looks good, though so far the images I've made don't look as nice as with 1.4, but I guess that's largely down to finding the right tweaks for the model and right magic wording for the prompts.