Stable Diffusion AI Image Generator

Automatic1111 Webui has now been updated to support Pytorch2 and some other tweaks and features. I'm not seeing much difference in rendering speed, but I was fine with it before, so no biggie. I had to manually (re)install Pytorch and Xformers to get it to work.
 
OnlyOneKenobi said:
Automatic1111 Webui has now been updated to support Pytorch2 and some other tweaks and features. I'm not seeing much difference in rendering speed, but I was fine with it before, so no biggie. I had to manually (re)install Pytorch and Xformers to get it to work.
Click to expand...
Apparently xformers gives no real benefit with Torch2?
You're meant to use --opt-sdp-attention instead.

github.com

Optimizations

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

Also are you getting a pip upgrade Notice whenever starting?
Pip is on the version it says is newer already, but it's still telling me to update it lol

Never mind, pip seems to have magically sorted itself lol
 
Vorastra said:
Apparently xformers gives no real benefit with Torch2?
You're meant to use --opt-sdp-attention instead.

github.com

Optimizations

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

Also are you getting a pip upgrade Notice whenever starting?
Pip is on the version it says is newer already, but it's still telling me to update it lol

Never mind, pip seems to have magically sorted itself lol
Click to expand...
Yep, in fact using Xformers in the new version caused issues whenever I was rendering any more than 3 images at a time.
Disabling it solved the problem. PIP still whines about being outdated, doesn't affect anything other than the complaining text on screen though, and only when installing or upgrading some components.
 
>goes to Civitai
>looks for highest-rated LoRAs.
>every single one is based on an underaged-looking anime girl

*sigh heavily*
 
Vorastra said:
>goes to Civitai
>looks for highest-rated LoRAs.
>every single one is based on an underaged-looking anime girl

*sigh heavily*
Click to expand...
Yep... found that if I want something artsy that SD can't do then I train it myself. There's a handful of nice styles on CIVIT that are nice, but for everyone one of those there's at least 10 000 waifu's or weird surreal porn crap.
 
Trying out Vladmantic Webui which is quite similar to Auto1111 but not broken at this point in time, unlike the latter which has various issues after the most recent update. It seems a little bit faster than Auto1111 but not significantly so.
 
OnlyOneKenobi said:
Trying out Vladmantic Webui which is quite similar to Auto1111 but not broken at this point in time, unlike the latter which has various issues after the most recent update. It seems a little bit faster than Auto1111 but not significantly so.
Click to expand...
What's broken? Haven't noticed anything.
 
Vorastra said:
What's broken? Haven't noticed anything.
Click to expand...
A few things, most seem to have been fixed though. Generating images intermittently hung, issues with certain LoRAs having no effect, issues sending stuff to and from IMG2IMG, inpainting, etc.
 
Finally got around to using ControlNet.
Very powerful. Finally I can get the poses and positions toward the camera that I want.
 
I asked ChatGPT to make a prompt on how it looks like:
It gave me:
Create an image of ChatGPT, a large language model, as a humanoid robot. ChatGPT has a sleek and modern design with a silver and blue color scheme. Its head is round and features a glowing blue orb in the center, which serves as its "brain." The body is tall and slender with segmented silver plates, allowing for maximum flexibility and movement. ChatGPT's arms and legs are long and streamlined with silver plating and blue accents, and each hand has four long and slender fingers. Overall, ChatGPT's appearance should convey a sense of intelligence, sophistication, and advanced technology.
Click to expand...

Bing create draws it like:

1683634285280.png

Anyone with access to Stable Diffusion or Midjourney that can give it a try, and post the result?
 

Attachments

  • 1683634200151.png
    1683634200151.png
    32.7 KB · Views: 6
Yeah, long winded creative writing style descriptions doesn't do much in the way of SD prompts. If you rework the core idea of it into the standard SD type prompt then you get better results but you'll almost never get exactly what you want by just prompting alone.

Without spending too much time or using any other models \ add-ons on this, this was made just with prompting and stock SD1.5, it almost matches the description except for the silver and blue colour scheme and number of fingers.

00013-2891354958-perfectly centered raw analog photo of a (robot_1.1) with a round head, (slee...png

The prompt / negative prompt / steps / sampler, seed, etc data hereunder.

perfectly centered raw analog photo of a (robot:1.1) with a round head, (sleek:1.2), tall slender body, modern design, silver plating and blue accents, high tech, studio backdrop, photorealistic, 8k, studio lighting, award winning art, trending on artstation, concept art by Greg Rutkowski and Alphonse Mucha, highly detailed, cinematic lighting, highest quality, masterpiece, UHD, HDR, creative, Particularly Detailed,octane render, trending on Artstation,unreal engine dreamy, Outstandingly Detailed, lumen, raytracing, subsurface scattering, 8k, cinematic lighting, masterpiece,3D art by David Levy
Negative prompt: bulky, combat, chunky, fat, man, woman, child, human, portrait, closeup, bad art, boring, blurred, unclear, speech bubbles, text, captions, subtitles, titles, writing, credits, artifacts, compression, mediocre, average, uninspiring, bland, flat, mundane, run of the mill, overexposure, cropped, watermark, logo, barcode, UI, signature, text, label, error, title, stickers, markings, speech bubbles, cropped, lowres, low quality, artifacts, bad_prompt, bad-hands-5, EasyNegative
Steps: 25, Sampler: DPM++ 2M Karras, CFG scale: 10, Seed: 2891354958, Size: 512x768, Model hash: cc6cb27103, Model: model
 
OnlyOneKenobi said:
Yeah, long winded creative writing style descriptions doesn't do much in the way of SD prompts. If you rework the core idea of it into the standard SD type prompt then you get better results but you'll almost never get exactly what you want by just prompting alone.

Without spending too much time or using any other models \ add-ons on this, this was made just with prompting and stock SD1.5, it almost matches the description except for the silver and blue colour scheme and number of fingers.

View attachment 1521463

The prompt / negative prompt / steps / sampler, seed, etc data hereunder.

perfectly centered raw analog photo of a (robot:1.1) with a round head, (sleek:1.2), tall slender body, modern design, silver plating and blue accents, high tech, studio backdrop, photorealistic, 8k, studio lighting, award winning art, trending on artstation, concept art by Greg Rutkowski and Alphonse Mucha, highly detailed, cinematic lighting, highest quality, masterpiece, UHD, HDR, creative, Particularly Detailed,octane render, trending on Artstation,unreal engine dreamy, Outstandingly Detailed, lumen, raytracing, subsurface scattering, 8k, cinematic lighting, masterpiece,3D art by David Levy
Negative prompt: bulky, combat, chunky, fat, man, woman, child, human, portrait, closeup, bad art, boring, blurred, unclear, speech bubbles, text, captions, subtitles, titles, writing, credits, artifacts, compression, mediocre, average, uninspiring, bland, flat, mundane, run of the mill, overexposure, cropped, watermark, logo, barcode, UI, signature, text, label, error, title, stickers, markings, speech bubbles, cropped, lowres, low quality, artifacts, bad_prompt, bad-hands-5, EasyNegative
Steps: 25, Sampler: DPM++ 2M Karras, CFG scale: 10, Seed: 2891354958, Size: 512x768, Model hash: cc6cb27103, Model: model
Click to expand...
Model: Realistic Vision V2.0 - Inpainting
Everything else the same.
00000-2891354958.png

00000-2891354958.png00002-2891354960.png00005-2891354963.png00004-2891354962.png
 
Last edited:
You must log in or register to reply here.
Back
Top
Sign up to the MyBroadband newsletter
X