-> PonyDiffusion / SD 1.5 | NoobAI V-Pred
Please, check the description, there is a little bit of useful information.
Indroduction:
This LoRA is supposed to represent Jadf's artstyle.
Please, be sure to use the same positive and negative quality prompt for better result.
Compatibility:
Humans: great
Anthros: great
Ponies: good
Multiple characters: good
Character LoRAs; good
Unique characters: good
Recommendations (Pony):
Trigger word - jadf (v2+)
1.0 emphasis
AutismMix - Pony for generations
score_9as quality tag, nothing moreold artstyle, flat colorin negatives
Recommendations (SD 1.5):
1.0 emphasis
(futanari, erection, size_penis, size_testicles:1.2) for futas
Size can be
large,hugeorgiganticWorks better with
hugeorgigantic
You can use these LoRAs for any penis you want:
Erect Horse Penis LoRA for erect horse penises
Flaccid Horse Penis LoRA for flaccid horse penises
Erect Human Penis for erect human penises
Large Flaccid LoRA for flaccid human penises
Characters (Pony v2+):
Jadf-tan
black hair, very long hair, blunt bangs, hime cut, red hairband, yellow eyes
There is still not so much images of her, but character is now working, at least. I hope you'll like this change, and of course, i hope there will be more art of her in the future.
Credits:
Trained on NAI Anime Full. Dataset of 212 pictures (v1), 20 epochs, Clip Skip - 2
Trained on PonyDiffusion v6 XL. Dataset of 360 pictures (v2), 20 epochs, Clip Skip - 1
Trained on PonyDiffusion v6 XL. Dataset of 600 pictures (v4). 30 epochs. Clip Skip - 1
Really big thanks to:
Jadf for beautiful artwork
Some good folks from Discord server
You, for reading, downloading and reviewing this LoRA
Description
Initial release
FAQ
Comments (26)
If EasyNegative doesn't work for this Lora, how about other negative embeddings like DeepNegative etc?
EasyNegative is not recommended by me for all style LoRAs you're using. Realistic or anime - doesn't matter. And there is only one reason. EasyNegative changes style of your generations. And you know what changes styles as well? Style LoRAs. Other embeddings like DeepNegative are doing very the same thing. In some different ways. So, if you don't like the result with your embeddings, just turn it off, style should become more consistent.
so I want to ask about your Lora training, could you please tell me about the value of your dim & alpha you're using? also about unet & text encoder value
Hi, for all my LoRAs right now i'm using 32 DIm and 16 alpha values. Talking about lr / unet / te values.
I'm using different values for diffrenet sizes on dataset.
For example, for style with 300-400 pictures in dataset i'm using 1 repeat, 20 epochs and 2e-4 lr / unet and 1e-4 te values.
For simple characters or concepts with dataset of 300-400 pictures, 1 repeat, 10 epochs i'm using 5e-4 or 4e-4 values for lr / unet and half of unet for te.
If my dataset is not large enought, i'm just rising repeat numbers, until i'll get 300-400 iterations per epoch.
You can check all my LoRAs settings in your WebUI LoRA tab. Just hover over any LoRA you want and you'll see small signt in the upper right corner. It would show you all LoRA's metadata. Training tags, settings, dataset size etc.
@TroubleDarkness Thank you, thank you a million times over, dear friend, for your incomparable act of goodness. May your kindness be rewarded tenfold, and may the universe shower you with blessings beyond measure.
@reinzharte Sure.
@TroubleDarkness can you do an update of all the values here? It's been 4 months, so I Guess you changed some things to improve training, right?
@DefaultBar yeah, sure, i've adjusted a lot of settings. I'll update Zheng's LoRA nad Jadf LoRA later. just don't have enough time to retag datasets.
@TroubleDarkness no I mean the values you replied this comment with, about the lora training, please I wanna see the new exact settings you're using now
@DefaultBar ah, this you mean.
For style LoRAs - Dim 128, Alpha 128
For character LoRAs - Dim 64, Alpha 32
For concept LoRAs - Dim 32, Alpha 16, sometimes higher
For 3000 steps (excluding batch size) - 2e-4 lr / unet lr, 1e-4 te lr. More steps - lower lr. Less steps - higher lr.
The same amount of epoch for styles, characters, concepts. Now i'm looking at amount of steps i need, depending on dataset, its quantity and quality.
For example, i have 100 images, i need 3000 steps. I'm just doing 30 epochs. If i have 300 images, i'm going for 10 epochs, but maybe a little bit higher for style LoRAs. As i said, all values i'm using depends on dataset.
You shouldn't use repeats for any type or steps increase. This meant to be used for balancing amount of images in subsets. So, 1 repeat everywhere, until you'll need subsets for training different clothes, for example.
@TroubleDarkness much obligued pal!
how much weight if i use lora erect horse penis?
I've used both LoRAs at "1" emphasis for testing, and got the same "good" results as with my other style LoRAs. Not sure if this would work on your model.
@TroubleDarkness Ok, thanks for the info, and thanks for all the Loras and the models you share 😉
Can I ask from what discord server did you receive help from? And if you could pass a link to enter?
This is just a tribute to those who helped me in the first attempts to create LoRA. Server is completely private.
Will you update this for Pony? This is one of my favorite styles
Same... I feel you. A month ago I started working on a PonyXL jadf lora myself with around 483 images. Now I only have to fix the auto-tagged images manually, but I'm lazy (progress is very slow). Every week I check if someone has done the job, so I don't have to do it. :)) Maybe in a month I'll finally finish my attempt on making one...
Hi, for some reason didn't got a notifications about your commentary. Right now i'm not planning to do something for PonyDIffusion-based models, because there is still a plenty to do with my SD 1.5 models. And i never worked on PonyDiffusionXL LoRAs, so, i have no idea what settings would be good for training characters, concepts and styles. It need a lot of testings. Most resources just misleading users about training settings (even with SD 1.5, still), so, you need to figure it out by yourself.
And still, original dataset needs to be expanded and adjusted, because i didn't know how to use buckets before. This release was 9 moths ago, i didn't know a lot those days. Maybe i'll try to gather new dataset for SD 1.5 LoRA update soon.
@SkagenM I'll stay tuned for the day :)
@S4NB Hi again, i've tried XL LoRA training yesterday and seems like i really can't train it with decent amount of steps. Takes freakin' a lot of time. Sadly. But anyway, i'll try to update SD 1.5 Jadf Style LoRA. I've already got new dataset, cropped and tagged. I hope it will be better than v1, but test version came out not so good as i've expected.
@TroubleDarkness I haven't given up on 1.5 models, I'll be sure to check out your new Lora when it comes out. Got any 1.5 Checkpoints you prefer over say Divineanimemix, HardcoreHentai and Meinahentai?
@S4NB Got it, anyway, figured out why test version of LoRA was looking bad, i forgot to set "Clip Skip" value to "2". Talking about models, the only SD 1.5 model i'm using is DivineEleganceMix. Right now i'm not working on Anime version. But i have no idea how other checkpoints are working right now, if it was updated.
Hi, i've trained a LoRA based on PonyDiffusionXL. Works pretty good with AutismMix, in my opinion. So, i'll release it soon. Previews already done, so i need to find some time to make a full release on CIvitAI.
@TroubleDarkness Hell yeah! I don't have to finish mine now :))
@SkagenM Updated.