All the preview images are raw outputs (no Adetailler/HiresFix)
Follow me on X if you want to be informed when a new model is published: https://x.com/AeremAi
It can produce nsfw/sfw content but you may need to use the correct tag in order to get sfw content without one of your character going around naked :p
Non exhaustive list of characters available with pony diffusion V6 XL:
https://docs.google.com/spreadsheets/d/1m2W-pZEvHuEpfHcNHrxCSr-Aw1mgtUUYho6sz9LChEA/edit#gid=0
(Spreadsheet by Marusame)
Description
FAQ
Comments (35)
differences in quality between the versions??? which one should we get?
Unfortunally yes, even tho it's the same lora just scaled down with kohyaSS :/ (Maybe I done it wrong, it's the first time I do this)
From my testing the "full" version is better. I added the smaller version for people who just wanna test the lora without downloading 800MB or can't do it.
@Aerem Yeah, it seems pretty obvious there will be a dramatic difference between 800mb and 56mb versions. That's like 256kbps lossless audio (FLAC or similar) vs a 32kbps MP3. What about one in the middle, around 250-400mb? Can you upload one in that range for a reasonable balance between quality and size?
@dita Sorry for the late reply. I don't find a good source of information about how to resize Lora with koyah_SS, so I will need to figure it out by myself. Unfortunately I don't have much free time right now, so I'll try but I don't promise anything.
@Aerem Okay, thanks. I appreciate it. I have 17GB free on my SD hard drive right now.
@dita Well for the moments I'm not very successful in my research, but you can download the V2 version which is only 7 mb. I hope you'll be satisfied with this version.
@Aerem Okay, thanks! I appreciate you trying.
only 7.18 megas?! how 'd you manage this madness?!?! good job btw it looks great, samples at least!
edit: if you have a bigger version that works even better i'd download that one =P if you upload it
For the 2.5DCartoon V2 and 2.5DRealistic versions, I merged two lora with different network dimension (2.5DCartoon (Dim:128) and a new 2.5D Lora more realistic (Dim: 32)) using the "super merger" extension in Automatic1111.
Default setting
Remake dimension: auto
And tadaaa ! A happy accident ! :D
I indeed have a bigger version, but the problem is that it work pretty great with anthro/furry characters but for human characters faces are quitte awfull most of the time. That why I merged it with the 2.5DCartoon version in the first place.
@Aerem i appreciate your work man. pony is so amazing, but it's lmost impossible to make a realistic style with it. Your work is gold.
do you add the lora to the adetailer prompt? i've experimented with anime chars, like chun li, and it goes a bit berserk with 2.5 realistic :P
Can you share your image of chun li or the prompt to see exactly what the problem is?
Otherwise, yes generally I add the lora to the adetailer prompt
@Aerem adding it now, yoiu'll see. it grabs realistic but derpy faced chun li (because she's not been adetailed yet) and it gives her a more anime face than the realistic one from the lora, not always, but sometimes.
i think its because the concept of chun li has been fully trained as an anime char
@Aerem posted it. compare chun li 3 pics vs 4th pic which is just korean woman - korean woman came out much better
sidenote: before adetailer chun li looked like a normal character pre-adetailing, some flaws to correct was all. I tried changing it from repeating the character name to writing "beautiful chinese woman" in the adetailer positive prompt and it works much better
@Malessar Thank you for your feedback. Yes, that could be it. Then the lora used in the base also has great difficulty with faces, so it may also come from there or a little of both.
I've published 3 images, using Adetailer with "Inpaint denoising strength" 0.75 in the Adetailer settings.
Image 1: no Adetailer prompt
Image 2: score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up,Chun Li
Image 3: score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up,Chun Li,<lora:Realistic000003_1_2.5DAnime_0.05_Merge1:0.8>
Images 2 and 3 looks better, but I don't know if they correspond to what you're looking for. Personnaly i'm not very satisfied with the results i'll try to improve the lora when I have time.
Please naming is important... This is very confusing :/ If I choose :
2.5DCartoon V2 I get -> Realistic_2.5DAnime_Merge1
2.5DRealistic I get -> Realistic000003_1_2.5DAnime_0.05_Merge1
2.5DCartoon I get -> 2.5DStylePonyDiffusionV6XL_V0.1
Smaller version I get -> LIGHT_2.5DStylePonyDiffusionV6XL_V0.1
Fixed for 2.5DCartoon V2, 2.5DRealistic and Smaller version. I'll fixed the 2.5DCartoon later.
Sorry for the trouble
Edit: 2.5DCartoon Fixed
Hi, Loving the 2.5D cartoon but the realistic one is bricking for me every single time.
Interesting, it like craps out hardcore at .7. Wasn't aware it would brick between .8 to .7
Hi, i'm glad to hear that you like the 2.5D cartoon style !
The 2.5D realistic really need some work. I work on the V2 right now, hopefully I get conclusive results soon. Thanks you for you feedback ;)
Did the current 2.5D realistic get pulled from use on-site here?
@coervus no, i think it got renamed. there's also a discrepancy with what the output file is called and the file name
@coervus It's should work again, sorry for that
thanks for that! Fragile as it is the results are amazing when it's working!
Could you please upload a version where the base model is SDXL 1.0? I am not able to import it on the image generator I am using.
Very good images but its like...so close, yet so far for me to run it.
Unfortunately, these lora only work with Pony Diffusion V6, so even if I could do it, it wouldn't work on the image generator you're using. Because I guess the problem is that Pony Diffusion V6 isn't available on the image generator you're using right ?
Thankfully, Pony Diffusion V6 is available on the image generator I am using. I imported it previously while it's base model still said "SDXL 1.0" , when the maker of Pony Diffusion V6 changed it temporarily due to import problems on other platforms.
Quick question, what are you setting for your instance and class when you are training these? I'm trying to train some myself but have been met with failures so far :(
I haven't used trigger word/instance tokken, and for the class I don't really use it, I always leave "ams" as default name. I hope I've answered your question.
In 99% of the case when training fail it's because of the dataset used.
Otherwise, if you want you can use CivitAI for training lora on Pony
Is training a lora for Pony just traning a SDXL lora using it as your base checkpoint, or does it require anything different as far as the process goes?
To train a lora on pony, simply use pony diffusion as the basic model. It might also be interesting to use special tags such as source_anime, rating:explicit and score in your dataset but that's just a guess.
Thank you for keeping furry capabilities in your models 🙏
You're welcome!
please explain the alternates???
Hi, It's just a name. They don't fit in with the other two main styles "Realistic" and "Cartoon", so I opted for the name "alternative". But since English isn't my mother tongue and I was really tired when I published these two models, I stupidly called them "alternate" by mistake. :')
Sorry for the confusion
Edit: I changed the name to "OtherStyle"
Details
Files
2.5DRealistic.safetensors
Mirrors
2.5DRealistic.safetensors
2.5DRealistic.safetensors
2.5DRealistic.safetensors
2.5DRealistic.safetensors
2.5DRealistic.safetensors
Realistic000003_1_2.5DAnime_0.05_Merge1.safetensors
2.5DRealistic.safetensors
2.5DRealistic.safetensors
hf_model_41.safetensors
2.5DRealistic.safetensors
Realistic000003_1_2.5DAnime_0.05_Merge1.safetensors
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.