you need the model information and the trigger words.Originally Posted by semthegreat
the specific model information is in the form of a LORA, Textual Inversion, or Lycoris and the trigger words are based on the specific model information.
Civtai.com shares user made model information for various models. take a look and go through the tutorials there. it will cover everything.
Last edited by darkbladeopiate; 20th July 2023 at 22:11.
Liked by 1 user: roger33
depends on your PC
i use Automatic1111 webgui with SD1.5. completely offline but it take a decent GPU i have 1080TI it works well enough but i have extra cooling in my pc.
for easy install check out SEAIT
https://github.com/diStyApps/seait
Liked by 1 user: roger33
I have both an EVGA 1080 ti/ftw3 w/16gb of RAM and a EVGA 3080 FTW3 w/10GB RAM in different computers [the 3080 is in with a i9-13900k and 64 GB RAM]. It takes MINIMUM 3x as long to do anything on my 1080 than my 3080, in some instances, the 1080 reports a day of processing needed for what the 3080 can do it it under 15 minutes. It doesn't always work out in that ratio, but, I wish good things upon you in the form of a card that can help you spit bigger things out, even faster.
If i had to go back in time to tell myself anything before buying a more current generation video card (I bought the 3080 after pandemic but before self hosted ai image generators were this mature).. it's buy one with more RAM! It looks like when SDXL fully drops, it can be a pig - in the sense some toolset developers are recommending 24GB RAM for a fine-tuning. 12GB for LoRA training - I really only have been experimenting with textual inversions, but they fly training them on the 10GB 3080 FTW3.
Maybe I should start uploading some of my stuff here, duh!
Liked by 1 user: roger33
well if you can deal with complexity comfyUI i believe is currently one of the best UIs for generation.
Liked by 1 user: roger33
Is there any way I can get an AI to learn an entire body/face/outfit of someone and show them from different angles?
Yes, you can train a LoRA. Lookup kohya_ss to know more
Liked by 1 user: roger33
I'm looking for a version that can take a facial image of someone I know and create a similar looking ai image in various porn images. Is there one that can do that?
Liked by 4 users: pixar, possiblethrowaway3, ppunter, roger33
I'm using Vast.ai. You cant rent GPU's cheap, they have an SD Automatic1111 docker image that works out of the box in their recommended images. Paying about $0.20 an hour, can generate 8 images at 512x512 and 30 steps in about 5-10s, there are no queues as it's your GPU for that time. You do have to download models each time you start an instance, unless you want to pay for persistent storage (it's cheap, I'm being cheaper). So make sure to find a server that has good upload/download speeds. I think you can also connect Google Drive etc, have not toyed with that. ALSO be sure on the Rent screen to increase the storage. It defaults to 16gb which is not enough if you're using a few models. I don't think this increases the cost, I usually did about 50gb and had plenty of space. You CANNOT change this once you rent (that I could find), so you'll have to destroy and re-rent if you forget.
To download models, you'll SSH in to the instance using Putty (there are tutorials that I think are easy to follow on Vast). Once you're in go to the models directory in the instance. Command is "cd ./workspace/" followed by the rest of directory. I usually cd one level at a time and I'm not looking at it now. Something like /workspace/Stable diffusion/models/stable diffusion/. Leave that for a sec
Go to Civitai (possibly others, haven't tested). Find the checkpoint you want, right click it's Download button and copy link. Go back to Putty and type in "wget **right click here to paste** --content-disposition" and it should start downloading in a few seconds. If you've picked a good server this should take 1-2m per 2g checkpoint. I copy link and paste into excel, then have a line that concatenates wget - copied link - --content disposition, because im lazy and don't want to type that much. Also don't have to dig back through civitai once I find models I like.
For Lora's, etc, same procedure as above just change the directory in putty before you enter the command (.../models/Lora). You can move files with the mv command if you mess up. From Instances in Vast.ai you can access Jupyter, this is an easy interface to pull your files from (/workspace/(SD...)/outputs You can also upload files here, but for models it was much too slow.
Once you've got that in place, proceed with Automatic1111 webui as normal. Click open from the Instance menu. I'm still learning here, but at $0.20 an hour I'm pretty good with that. Be sure to pull your files and then also be sure to destroy the instance when done, otherwise you will keep racking up time. I'm not sure but I think it charges by hour, so if you only use 5 min that's an house, if you do 1hr and 5m thats two hours. Let me know if you try it out and if this was helpful!
Here is how to run Ai Image Generation on your own computer. You get privacy and no subscription costs.
You need a reasonably decent computer with at least 16Gb RAM and a decent video card. At least one with 8Gb of video Ram (VRAM), although Auto 1111 allows you to work with 4Gb, too, but it will be slow. You can even run with just your CPU, but it will be horribly slow. If you want a great and affordable Win computer get a refurbished HP Z Workstation, add an nVIDIA video card (RTX 2XXX, 3XXX, or 4XXX). Be aware that the faster RTX cards use a lot of power so you need at least a 700 watt power supply, which has 6, and or 8 pin power connectors. You can get power connector adapters that let you hook 2 4 pin connectors together to make an 8 pin connector. If this is all greek to you, find someone who really knows computer hardware. Game computer players who know hardware will know all of this about Video Cards because they use powerful ones.
Check these out. EasyDiffusion is, in my opinion, somewhat easier to learn for beginners, otherwise the two one click installs are pretty similar
Best Video Tutorial to Install Automatic 1111 (easy install that runs Stable Diffusion)
Monzon Media
https://youtu.be/A2s9OnGacxA
Best Video Tutorial to Install EayDiffusion v 2.5 (easy install that runs Stable Diffusion)
Monzon Media
https://youtu.be/gJeFxd1O_90
No one explains this stuff as well as Monzon Media.
These videos are for Windows computers, but you should be able to find similar, probably not as good, tutorials for Mac and Linux.
Enjoy
Liked by 4 users: deepsepia, Jan_Klaassen, roger33, xbasilx
I've been alternating between https://dezgo.com/ https://www.anydream.xyz/create
https://www.craiyon.com/
https://pornderful.ai/
https://pornpen.art/
Cheers!
Many thanks for these. Checked out dezgo, anydream & craylon, and so far anydream fits my needs for injecting images to generate images.
Liked by 1 user: roger33
Hi,
.
i've a few questions -
1) Can we use sample particular face , exact body shape/type from bunch of many images , and thus train ai to learn them , & so create new images based on them.
2) if you say for LoRA thing , Can kohya_ss work with Automatic1111.
3) Also can we use multipe extensions together , like controlnet + some realistic ones over existing or newly created images , to create further enhance modifications.
Liked by 1 user: roger33
I used https://playgroundai.com/. It is probably not so powerful but very easy to use
Last edited by sophie_bas_nylons; 31st August 2023 at 11:09.
Stockings for all
Liked by 1 user: roger33
Altering genitals would be nice. Not just ladies with penises or with two penises, but smooth, shaven ladies with no vulva and men with no penis or no testicles.
Liked by 1 user: roger33
that is specifically what Lora's are for. they are specifically trained data. and as far as i know any Lora work with Automatic1111. Textural embeddings and hypernetworks are also used. as well as lycoris and locons. all work with automatic1111. and yes you can use multiple extensions. not sure what you mean by some realistic ones.
if you want to get in depth and learn specifically whats being used to create your image then learning comgyUI would be something to go for
automatic1111 was updated recently to now work with SDXL 1.0. previously only comfyui was working well with it.
Posting Permissions
Reply With Quote