Stablegram
my fanslora : fanslora.com/miaou
US/European instagram woman style
Warning: Add "nude, naked" to the negative prompt if want to avoid NSFW.
V1-V2.2 : (Chilloutmix base) Trained with +70 000 pictures on EveryDream2
V2.3 : BasilMix base Trained with +70 000 Pictures merged with PoV Skin Texture - r34 Lucid Black + PoV Skin Texture - Dreamlike r34
Prompting: Simple prompts are better. Large amounts of quality tags and negative prompts can have negative effects.
Currently working on a new version with better accuracy and consistency. By the way, have fun and don't hesitate to share your images !
Description
VAE-FT-MSE-840000 added to de model
FAQ
Comments (16)
Can it do male subjects and general concepts?
Of course, I trained the images with chilloutmix so yes you can do everything
Wow, 70K images is a lot.
It must have been very challenging to conduct the training, I salute you for the effort.
I'm working on my own realistic training, it would be very helpful if you could share your experience with training parameters that you found to work best for you, like resolution, batch size, learning rate, etc.
You just shared the huggingface link.
The chilloutmix.ckpt that you have on huggingface starts with hash "249039c5", is the one that you used for your training? Is it a convention that you did to Chilloutmix-Ni-pruned-fp32-fix from safetensors version? Have you added 840000 VAE?
You shared here epoch 2, and on huggingface you have epochs 3-5, do you plan to continue the training with more epochs?
I imagine it's taking a long time...
Can you please explain what's the purpose of the training, to improve the aesthetic, change the base style of Chiloutmix, or something else?
You haven't specified trigger words, are there any specific ones that you have added? Can the model produce different European nationalities better? Can it produce certain named people better?
Hi Alex,
a have used Chilloutmix-Ni-pruned-fp32-fix.safetensors, converted to ckpt to be able to use it on EveryDream(it seems to me that it is possible to use a file safetensors directly but in my case I did not succeed), no VAE as been added. I don't know if charging a vae from stable diffusion would be the same thing (?)
i used Runpod with A6000 48GB Vram GPU, LR 1e-7, 20 batch size
in reality it is the epochs 9/100, it's more readable for users, I'll update it like this v1, v2, v3...
it took me 14 hours for example to get the 9th epochs, so yes it takes a lot of time!
all the images of women, for example are trained on the trigger word "a woman". so a woman trigger word gives very different results for each result there can be a black woman, white woman or even asian woman with always an instagram woman style
for the moment the training is on break, I will change all the captions. for is model I use the basic model clip comes with stable diffusion, it's a big mistake for me I will redo everything with ViT-L-14/openai
the purpose of this training is to obtain images like https://www.instagram.com/callmybunny/
"instagram girls style" there are a lot of Asian models on civitai but I couldn't find a US/European one so I decided to create it
i used Chilloutmix as a base because I think it's the best in terms of photorealism
Chilloutmix is very nsfw oriented, so my model too... it's a us/european chilloutmix. I did some nsfw test it works very well but for logical reasons I would not share this part... it's training on real girl & man is can be very offensive for the real people behind these images...
last thing, on 70k image there is a lot of images of men, landscape, shoes, clothes, so I trained a bit of everything!
@Shierry
Thank you very much for all the information.
Do you plan actually go up to 100 epochs?
@Shierry
I want to invite you to my 🫶 Discord server.
I need training buddies... 😊
@alexds9 i thing I'm going to start all over again with another captions model (ViT-L-14/openai)
This is crazy!! these gens are just woah, you're going in god tier levels of training and model making, don't bake a vae into the model i hear it can damage quality and accuracy, im gonna wait for v2 as you said your currently planning an update, but this is by far the most insane model iv seen 70k images, i really hope it'll work with celebrity faces, im looking forward to this big time!! Keep up the great work and i may even throw a donation your way
Hello, thank you for your comment.
a new version is coming tonight, it will be totally different and better
What do you use? Sampling method? upscaler?
TLDR: Could someone post a vagina generation pic of this model? Just want to see how good it does.
This model any good at nudes? particularly vagina detail. Most realistic models do poorly with them. Litearlly most realistic models, its just a simple black line drawn "down there". My internet is slow, so just asking so i dont waste time downloading TY. The only realistic model ive found that does decent default vaginas is lazymix, nothing else comes anywhere close. I hate using vagina LORAs too, they're very tedius and annoying to use. Rather a model just have good decent defaults to begin with.
EDIT: I folded and downloaded it. Im not an artist, pro photographer, or graphics designer or anything, just a casual pleb that uses use stable diffusion for porn. Not really interested in SFW content, so as ckpoint models go, this one id rate 3.5/5. Its ok, but I'll stick with lazymix for now, its better at NSFW stuff.
im gonna try for you