Use 'sw33ny' to activate the lora.
There are two versions available, Base Pony and Pony Realism.
Pony Realism is trained on Pony Realism and would give best results on realistic models.
The second version is trained on vanilla Pony and is not as accurate but works on all pony models.
If you enjoy my work please leave a like or a review. Pics of your work are also appreciated :)
Description
Trained Pony Realism
Use this for realistic pony model like Pony Realism, everclear and zonky.
I found the most realistic results are with DPM++ 2S a Karras though most samplers such would give good enough results.
FAQ
Comments (15)
Very good likeness all the time, you are doing a great work.
This is really really good. I am trying to train LORA's myself, and am not quite getting to where I want to be. On the same model as this specifically. Could I ask you a couple questions.
1) Approximately how many images did you use for training this one?
2) Are you using default parameters ? If not, would you be willing to share your parameters for this one, so I could check them against mine and see if I am missing something obvious. I have read article after article, and am getting results that I'm like 80% satisfied with, so just searching for the other 20%.
Cheers
Mostly default params with training batch 2. Trained on about 140 images. I used wd14 to tag then pruned, fixed and added tags on top of that. For realistic people I found training on Pony Realism helps a lot with accuracy but it limits on the amount of checkpoints you can use the lora on. For pony realism I find around ~2200 steps gives a good result. For base pony somewhere between 2200-3000.
Honestly I was also about 80% happy with the results on Pony and training on Pony Realism gave me the results I was looking for.
Pony realism is what I used on my most recent, but with more like 36 images. I used Civit's autotagger, which may also be part of the issue. I see so much contradictory advice on tagging/captions. Step wise, I'm right there....I suspect my issue may be too few images. Thank you for replying so quickly! Off to build a bigger cache of images!
@kestinek 36 Can be enough depending on the images themselves. The civitai autotagger is pretty good just make sure to also do a pass on the tags to make sure they are correct or if some need to be added to the pic. When tagging make sure you dont add words that you want to be associated with the trigger word. If you are doing of a person, remove words like mouth, eyes, eye color, moles etc unless they are unusual for that person in that case leave them. Basically you tag what you dont want to be in the lora so what you do want in it will be associated with the trigger word.
That images themselves need to be diverse with the subject of the lora being the constant. So same person in different clothes, poses and backgrounds.
@JaneB Largely that's what I've done, but I am going to take a hard look at my tags and images and see if there is room for improvement there. Thanks again!
@JaneB Going back over captions after your advice, I'm seeing where I did leave a lot of things in like "teeth" that could be causing issues. But that leads to another question, I have tags I'm not sure what to do with "1girl", "Solo", "looking at viewer"...If I'm reading your tip correctly, I leave those tagged, so the training won't spend time learning those. Correct?
@kestinek 1girl and solo you can probably remove because they are extremely generic, as for "looking at viewer" you should keep as it does describe the picture but not the concept (unless you are making a lora of looking at viewer)
@JaneB Hey, I just wanted to say, with your coaching here, I finally produced a character lora I am happy with. And seeing what has worked and hasn't worked, makes me excited to try some more. THANK YOU!
I'd love to give you props in the description, but since I am going to use the lora, to make both SFW and NSFW, I don't want to presume you're comfortable with that.
@kestinek You don't have to, I'm always happy to help but if you want I don't mind as long as it's something not too extreme
@kestinek Glad to see your success story here and I can see the lora on your profile itself looks good. I've been struggling to get good results with PonyRealism after seeing loras like these. I can follow the tagging changes you made but did you also increase your image count > 36 or did you stick with the original number and just change the tags?
@JaneB If you don't mind answering a few of my follow up questions to you: do you use regularization images? Do you have a target loss value? And did you do 2,196 steps per epoch or 2,196 steps total?
I'm guessing 2,196 steps per epoch, for a total of 39,528 steps, because otherwise you would be doing less than 1 step per image with 140 images. Either that or you are using some other process I'm not aware of. I've personally always used kohya_ss but if you are using a different process, that would be useful information as well.
@Clocksmith Sorry for late reply, no I stuck with 36 images....BTW, if I can help, feel free to dm me
This lora is great and with really great likeness! Hope you do more celeb pony loras! You deserve buzz for this.
Thank you! Got a few planned out though can't commit on when they will come out, got distracted with a few other loras