KNK SeryalPony
Is a merge using Pony V6 as base, with the porpuse of get a semi-realistic style (like my SD1.5 model KNK Semirealistic Merge).
I tried to get a Lora to get it but it doesn't work at all..., for that using the super meger extension adding lora weights a little by little until I got something reasonable.
Works well with almost any Lora created for Pony/Autismmix,
No one image in the previews was upscaled or treated more than increase/reduce the sampling steps
Recommends
Prompt:
score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_upNegative Prompt:
score_1,score_2,score_3,score_4,score_5,score_6, source_pony, source_cartoon, source_furry, monochromeI know, the tags are overwrited in low levels. But for some reason, this is the config that gives me better results.
CFG Scale: 6.5
Sampling Steps: 20 - 30
If you got deformed characters could be the resolution that are you using, Pony (and the base SDXL model) was trained in this resolutions, try use them:
(1024, 1024),
(1152, 896),
(896, 1152),
(1216, 832),
(832, 1216),
(1344, 768),
(768, 1344),
(1536, 640),
(640, 1536)Remember, your reviews and generations are helpful in guiding improvements. Feel free to leave your comments or requests, but I can't promise anything.
License
This model follows the same Pony V6 License.
Description
FAQ
Comments (4)
Can anyone please explain what's the point of merging loras into a model and publishing it as a new one? Merging several models with different weights after experimentation sounds reasonable, you don't need all those models in the end. With loras you could theoretically have a list of loras and their weight combination you find interesting and others can reproduce it on their end, then maybe experiment a little and find another combination to their liking. When loras are merged in you lose this flexibility. And there're a lot of merges done like that. Finetuning can add new concepts or divert the base model to another direction, merging rarely does that. I've no training experience, so maybe I'm missing something and such merges are useful in some context.
I could talk in my experience. I'll try to be concise, I prefer use loras to add features to a model for the next reasons:
- Is faster and easier control a merge than use fine tunning. If any feature is over-learned, I just can reduce que lora weight and retry in just 1 or 2 minutes. In finetunning, that could take hours... or days...
- I don't need trainning data for concepts that already have a lora: I just could check loras with an open license to use in my merges
- Confort: Yes is possible just share the lora recipe, but (for me) is easier just have all in one, and just now, the XL loras are around 40-200MB.
- Have a lora all-in-one? I tried, but for the diff in dims/alpha and type (Lora,Locon,Lycorys), they don't look anything like they do when they're separate.
That is my point of view, maybe the first one is my main reason.
Just for clarify, I didn't use any other xl model in this merge, just was lora merges
@Konoko Thank you for the thorough clarification. I agree with the main argument, merges with no doubt are much much faster, in this case it'll overshadow any of contra-argument. The best I can come up with is that when you use loras you need to be sure it does what it declares in many scenarios, which means a lot of prior testing before taking it into the merge, it also takes extra time. While in fine tuning you prepare your own data and there's no dependency on external bits (that of course is another complex time consuming matter). I can agree that merging lora is better than using them separately only if the number of them is large, did you track how many loras you merged in? And more importantly what new did it bring comparing to other models, e.g. autismMix, maybe new concepts/styles?
In any case, I'm intrigued and will give this model a try.
@green_anger You're right, using third party loras I lost the posibility of control the dataset used. In this merge I didn't add any concepts or characters. I just added several styles trying to stabilize the style and make it looks like a 1.5 model that I did time ago.
The lora's quantity wasn't the problem, the weight of each one was. That's because when you make the merge, the result isn't exactly as the preview, and you have to re-merge again certain loras (this was try and fail) trying to fix it.
Actually my old model (1.5 model) contains a lot of Hiten works, but I didn't find a Pony Lora with that style. I'm just creating it to improve this merge and share it for use stand alone. (I hope publish it tomorrow)