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

  • 🐕 I am attempting to get the site runnning as fast as possible. If you are experiencing slow page load times, please report it.
I cannot recommend Forge or any fork of A1111 over ComfyUI unless you like to tinker around with things on your phone or iPad while being connected to your local network.

Paired with LobeHub, which is a front-end overhaul for A1111/Forge, it’s just fun to play with. Forge has lots of tweaks on the backend that come out-of-the-box which makes it much more efficient and streamlined than a typical A1111 install. Things like the Kohya HiRes fix, ControlBet, HyperTile, etc. are all included natively (without any real explanation to newcomers as to what they do).

I digress, if you’re using Comfy, you probably like to have that granular control over your generations and Forge will offer you very little unless you just want to play around with SD while lounging around.
I mean, I don't not want to lounge around and use my ComfyUI install from anywhere, but it's not a burning need. ComfyUI has a --listen option, iirc so I figured I could probably just expose the relevant port from my desktop to my home network and use it from anywhere if I wanted. But maybe I'll look at Forge.

Thanks!
 
  • DRINK!
Reactions: Looney Troons
I made the switch over to Forge (created by the guy who pioneered ControlNet) from A1111 a few months ago. Performance is pretty great across the board.
I tried SDNext, Forge and Comfy. SDNext is overhyped garbage that shits the bed if you install an extension, to the point where even if you disable the extension you have to nuke your own install.
Forge was actually fucking slower than Auto1111 (AMD user, so that might be why.)
Comfy was too autistic for me, but it looks good.
Auto1111 has been reliable for me, even though I wish it had better performance. (Hiresfix dies on me, not enough VRAM)
 
I tried SDNext, Forge and Comfy. SDNext is overhyped garbage that shits the bed if you install an extension, to the point where even if you disable the extension you have to nuke your own install.
Forge was actually fucking slower than Auto1111 (AMD user, so that might be why.)
Comfy was too autistic for me, but it looks good.
Auto1111 has been reliable for me, even though I wish it had better performance. (Hiresfix dies on me, not enough VRAM)
I definitely agree about SDNext. I had nothing but issues with it when I was tinkering, which prompted me to try Forge out. I’ve got a fairly stout computer with a 4090 and was honestly surprised to see performance gains. Using the Kohya Hires Fix, I could generate 1080p images using the DPM++ 3M Karras sample at ~120-140 steps while employing Adetailer (~40 steps) and get an excellent image generated in about ~50 seconds. A standard A1111 install could yield similar results in about 1 minute and 10 seconds. Not super huge gains, but definitely noticeable if generating images in batches.

If you’re not against delving back into the autistic spaghetti that is Comfy, I stumbled across this workflow recently while trying to help a friend out. It’s a great starting point that will get you on your feet with generating some crisp images and it’s fairly easy to use. The instructions provided are definitely written by someone who isn’t a native English speaker, but what matters most of all is that the workflow is sane and makes sense. Snag the version 1.0 .json.

Be sure to also download the Comfy Manager. Comfy Manager has this awesome feature that allows you to load up a workflow (like the one listed above), and then identify and allow you to safely download any of the extensions that you might be missing to use that workflow.
 
Still an AUTOMATIC1111 scrub who recently updated to the newest version. I'm contemplating on moving on to SDXL, which I hear is easy enough to use as new versions of AUTO will immediately recognise it. Anything I should know before trying XL out like changing arguments to my .bat or deleting folders like /venv/? My GPU is an RTX 3070 so I'll probably take it easy and only do a batch size of 1 per generation.
 
Still an AUTOMATIC1111 scrub who recently updated to the newest version. I'm contemplating on moving on to SDXL, which I hear is easy enough to use as new versions of AUTO will immediately recognise it. Anything I should know before trying XL out like changing arguments to my .bat or deleting folders like /venv/? My GPU is an RTX 3070 so I'll probably take it easy and only do a batch size of 1 per generation.
It's plug and play. Drop SDXL and SDXL refiner into the models folder and restart your AUTO, it should work out of the box.
Some starting advice, do 50 steps and switch to the refiner at 0.8. Switching to the refiner earlier tends to make pictures "too detailed", while switching to it later tends to make it omit small details. For example generating a city without the refiner will make it generate very flat surfaces devoid of detail, generating with a reasonable refiner will give the buildings details like individual windows, rooftop antennae, parked cars etc, and too much refiner will add windows in places that don't make sense, multiple antennae, and add lines or patterns extending out from other things, in a way that never looks good.
 
  • Informative
Reactions: Baraadmirer
Are the tutorials in the OP still relevant/up to date? Figure I should ask before I spend a few hours installing one just to shitpost.
 
a coomer who jerks off to pictures of anime girls

Close, but missing the mark

1716626118741.png
 
Mob rule has dictated you are to be removed from the thread.
The funny part is when people talked about your first class trolling I was expecting way way worse than what you actually are generating.
It gets worse better. I'm still a horrific amateur, honestly.
15-2.png

8-2.png

7-2.png

5-3.png

3-3.png

Gas is hard on SD.
 
Last edited:
Back