A merge of SweetMint and some of my own loras. It's still a WIP but it's usable.
The generation settings i use:
Sampling Method: DPM++ 2M Karras
Sampling Steps: 25
CFG Scale: 9
Width: 768
Height: 1024
Hires Fix: ESRGAN_4x
Upscale by: 2.5
Works mostly fine without Hires Denoiser but gives better results for me if set to 0.2
I tried to refine the face generation into a slightly different direction since SweetMint was too similiar to other Checkpoints and i aim to get a more unique face style.
Description
FAQ
Comments (20)
reading "sour mint" as name instantly reminded me of a fun video about "mentos". old internet freaks of the 90's like me should know which video i mean....
Oooo-oooo, Ah-aaah!
How come you stripped the generation parameters from images and manually edited yourself? When trying to copy parameters from the model preview in AUTOMATIC1111/stable-diffusion-webui there are missing parameters...
i have this issue where the parameters dont get uploaded with the pictures and i dunno how to fix that, everything is up to date on my end so i thought it might be just civitai messing that up. kinda sucks cuz it takes a lot of time to manually get the params out of the img details and paste them 1by1...
@Notstable Are you using Google Chrome?
@ishimarukaito565 yes
@Notstable And you save/upload generated images by dragging or right click copy I presume?
Chrome strips exif data which is where the parameters are stored if you use copy from context menu or dragging. It is a preventative measure to avoid png bombs which are malicious png files. You should use your file explorer to upload and if you are not doing local generating of images but online I suggest switching to Edge. Much better and lower performance imapct over all.
@ishimarukaito565 i actually use the browse in FileExplorer way and i dont really like Edge that much. Performance isn't an issue to me since i'm running a i9 13900KS, RTX4090 and 64GB Ram :D
I saw some extention for A1111 to CivitAi, i will give it a try later!
@Notstable "png bombs"? How they wors?
@settima_ai it is just malicious packets disguised within PNG that when opened could for example self replicate exponentially causing device to oom in no time. There are several other implementations but not gonna go through it all here
@ishimarukaito565 I would like to understand how can duplicate if png is not executable code. But I cant find any reference to "png bombs" on internet, may be they are known with another official name?
@settima_ai A file does not have to be an executable in order to malicious. But here. I found one reference to it. You are unlikely to find much info regarding this stuff unless you regularly spend time in Security Research forums/boards/communities. The technique of performing a DoS attack by sending a PNG bomb is presented (opennet.ru)
@Notstable Can you tell how much seconds it takes to generate 1 image like in you demo gallery?
I'm using Google Colab Online and what to know if a real 4090 would be faster.
@settima_ai I can probably answer that. Yes it will most likely be faster. I am assuming you are using the T4 GPU on Colab. But remember that buying a 4090 without having a system that will keep up is just a waste. the i9 is not necessary but I would suggest at least go with the i5 13600K, minimum 64 GB DDR5 and do not cheap out on Motherboard since they are the one part that tends to give up first if you buy cheap. Also make sure your Power Supply is minimum 80 plus bronze rated and no less than 850 W. 1000 is preferable.
@settima_ai it takes a little more than 1 minute. the last pic i generated took 1 min. 2.2 sec.
and @ishimarukaito565 is right, there is no point in having a 4090 with a mid tier system. for stable diffusion alone, the i9 would be too op. if you are doing video editing, 3d-rendering and streaming like me, the i9 is a beast. and i also agree with the 64GB ram, my average usage is around 30's when im in my workflow so having the double of that is pretty neat to not be limited
@Notstable Infact I'm planning to buy a complete new PC. If all is pumped up as you suggest, the price is around 3000 Euro. But there is a doubt: the times you listed are similiar to what I'm experiencing using the T4 GPU on Colab. 1 minute and few seconds more... I was hoping you replied just few seconds...
@Notstable i5-7600k, 32GB RAM, 4060 ti 16GB, 34 seconds for generating 768x1152 and hiresfix to 1152x1728. You sure you have your optimizers properly set up?
@Notstable Here is another one. Not hiresfix but directly generated 1160x1624 in 57 seconds
https://i.imgur.com/bwqYS2n.png
@ishimarukaito565 here u see what settings i use: https://gyazo.com/0b3287819b0ac0891e1df3cc16ced67a so having a generation base of 768x1024 and high res around 2.5x-3.0x with 0.2 denoising just takes a bit. i think if i dont use the denoiser and lower the high res it will be faster ofc but i really need high quality images and that denoiser really makes everything look better for me. also i use xformers, i heard that i am supposed to use that one, what do you think? i also heard there is another one but not as good as xformers?
@Notstable Well I see a bunch of issues. Did you know that you have decimal weights that is not closed within parentheses in prompt? And that you decimal weights are far off scale from what is actually needed. In fact. Your prompt is pretty much having a civil war fighting itself. I have some stuff that can help. Also I generate directly at 960x1280 with SD 1.5 and then hirefix to 1440x1920. hehe. Special model. Add me on Discord, https://discordapp.com/users/215928417239957504