Stable Diffusion, NovelAI, Machine Learning Art - AI art generation discussion and image dump

[Ravier Xavenblood]

Do you have a decent GPU?

Is an RTX 4060ti 16GB decent in 2024?

 

[macrodegenerate]

Is an RTX 4060ti 16GB decent in 2024?

It's alright for stable diffusion.

Follow the instructions on this page:

GitHub - AUTOMATIC1111/stable-diffusion-webui: Stable Diffusion web UI

Stable Diffusion web UI. Contribute to AUTOMATIC1111/stable-diffusion-webui development by creating an account on GitHub.

github.com

Others will recommend comfy-ui, but this is the fastest way into the game for noobs.

Download SDXL and throw it into the stable-diffusion directory in models.

https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0

Alternatively if you want something with danbooru style tags.

Go to civit.ai and download either PonyXL or Noobs/Illustrious.

Edit:ninjaed.

 

[Baraadmirer]

It's alright for stable diffusion.

Follow the instructions on this page:

GitHub - AUTOMATIC1111/stable-diffusion-webui: Stable Diffusion web UI

Stable Diffusion web UI. Contribute to AUTOMATIC1111/stable-diffusion-webui development by creating an account on GitHub.

github.com

Others will recommend comfy-ui, but this is the fastest way into the game for noobs.

Download SDXL and throw it into the stable-diffusion directory in models.

https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0

Alternatively if you want something with danbooru style tags.

Go to civit.ai and download either PonyXL or Noobs/Illustrious.

Edit:ninjaed.

@Agares I'm going to disagree with macrodegenerate here and suggest the Forge UI instead if you don't feel like using extensions. It's more optimised and will take a shorter time to generate images, not to mention virtually no latency if you're switching from one checkpoint to another, as well as quickly changing settings like clip skip or VAEs.

 

[Catgirls are Love]

I'm going to disagree with macrodegenerate here and suggest the Forge UI instead if you don't feel like using extensions. It's more optimised and will take a shorter time to generate images, not to mention virtually no latency if you're switching from one checkpoint to another, as well as quickly changing settings like clip skip or VAEs.

It's also far easier to setup than Auto1111 which is really the best for anyone not entirely comfortable with git (ignoring that Auto1111 isn't being updated anymore).

 

[macrodegenerate]

The issue with forge is that there are several plugins which aren't supported that are really important. For example, regional prompting isn't supported as a plugin. That makes running the llama.cpp regional prompter impossible. That's a problem because as we move into the Flux era we are going to need regional separation. Even if flux is good at separating concepts.

 

[Baraadmirer]

The issue with forge is that there are several plugins which aren't supported that are really important. For example, regional prompting isn't supported as a plugin. That makes running the llama.cpp regional prompter impossible. That's a problem because as we move into the Flux era we are going to need regional separation. Even if flux is good at separating concepts.

There's a Forge extension (though I haven't tried it yet) called Forge Couple that comes pretty close to AUTO1111's regional prompting.

 

[Spruce Supremacist]

How's the AMD GPU support? Have there been any breakthroughs or is it still a big hit and miss on Windows?

 

[Overly Serious]

How's the AMD GPU support? Have there been any breakthroughs or is it still a big hit and miss on Windows?

It depends what your expectations are. You've been able to use DirectML for Stable Diffusion on Windows for some time now and if you're just doing SD1.5 or SDXL it's adequate-ish. But really you still need Linux and the thing you need to watch is ROCm support:

https://rocm.docs.amd.com/projects/install-on-windows/en/latest/conceptual/release-versioning.html

Same hardware but booted into Linux and my times were far ahead of using Windows. If and when we see ROCm 6+ support on Windows then you should start to see something closer to parity. But until then, and unless just having a little bit of a play with older models, I'd make yourself a Linux USB installation and boot into that instead. It's night and day, I'm afraid.

EDIT: Wait - that says 6.1 is now available for Windows? Okay, that's new since I last checked in. Maybe things have changed. Worth another go then, perhaps.

 

[macrodegenerate]

There's a Forge extension (though I haven't tried it yet) called Forge Couple that comes pretty close to AUTO1111's regional prompting.

Looks like I'm using forge from now on. That extension is great.

 

[Post Reply]

EDIT: Wait - that says 6.1 is now available for Windows? Okay, that's new since I last checked in. Maybe things have changed. Worth another go then, perhaps.

Pytorch doesn't support ROCm on Windows yet. The options for Win/AMD are still limited to DirectML, ROCm in WSL, and more niche options like Zluda and ONNX.

 

[Catgirls are Love]

There's a Forge extension (though I haven't tried it yet) called Forge Couple that comes pretty close to AUTO1111's regional prompting.

I've only done a few tests and experiments thus far (with a model I'm new to using) but it looks like it's a pretty good equivalent to Auto1111's regional prompter; it's definitely worth picking up if you're using Forge.

 

[BirdUp]

NovelAI has a preview of a new anime model out. It looks pretty damn good.






It can understand English now, so you don't have to use Danbooru tags in prompts, and you can have separate prompts for six different characters in one image. NovelAI's large language models are ass, though, so just buy credits instead of a subscription.

 

[crowabunga]

<snip>

I checked up on vg's /aids/ thread to see how they're taking it:

 

[AmpleApricots]

I got around to play a little with image models in the last few days. I need to change my default half-orc test to probe the models understanding because it's not difficult enough anymore. As aren't artstyles. Controlnets also are growing more powerful by the minute. I wouldn't want to be an artist. Next on my agenda is testing the Blender integration.

 

[Agent of Z.O.G.]

Can't believe how bad Dalle has gotten, barely better than the common free SD versions, and I think that's mostly because the promps don't carry over well (dalle is still better at having characters hold guns, for now).

 

[The Mass Shooter Ron Soye]

Pytorch doesn't support ROCm on Windows yet. The options for Win/AMD are still limited to DirectML, ROCm in WSL, and more niche options like Zluda and ONNX.

Here's your ZLUDA news, bro:
ZLUDA v4 Released For Initial CUDA Support On Non-NVIDIA GPUs (archive)

 

[bifftango]

Got back into generating images locally after building a new computer, so I thought I'd drop some of the best here.

More Sci-Fi book cover art:

Spoiler: Book it!

The Predator in sumi-e style:

Spoiler: Ink it!

Cheesecake postcard art, I guess:

Spoiler: Mail it!

And some attempts at an '80s ski comedy poster starring Pinhead, which didn't quite work due to Stable Diffusion not being able to generate readable text and skis on feet:

Spoiler: Jump it!

 

[Baraadmirer]

Illustrious (and to a lesser extent, NoobXL) seems to be a competitor for PonyXL now. Anyone try it out yet?

 

[Catgirls are Love]

Illustrious (and to a lesser extent, NoobXL) seems to be a competitor for PonyXL now. Anyone try it out yet?

I've used NoobXL a bit as I was trying for PonyXL alternatives that weren't quite as furry, and while I also have illustrious there's an Illustrious mix (NAI-NSFW-illustrious v7) that I prefer and I've gotten really good results from. Only downside is it's a very anime model so if you want some other style it probably won't do it well. My trend of using NSFW-trained models to make non-NSFW stuff seems to be continuing uninterrupted.

 

[Baraadmirer]

I've used NoobXL a bit as I was trying for PonyXL alternatives that weren't quite as furry, and while I also have illustrious there's an Illustrious mix (NAI-NSFW-illustrious v7) that I prefer and I've gotten really good results from. Only downside is it's a very anime model so if you want some other style it probably won't do it well. My trend of using NSFW-trained models to make non-NSFW stuff seems to be continuing uninterrupted.

Were you looking at roughly the same generation times for Noob and Illustrious when compared to Pony?