(optional)SDXL Recommended Resolutions/setting
640 x 1536 (5:12)/768 x 1344 (4:7)/832 x 1216 (13:19)/896 x 1152 (7:9)/1024 x 1024 (1:1)/1152 x 896 (9:7)/1216 x 832 (19:13)/1344 x 768 (7:4)/1536 x 640 (12:5)
(optional)SDXL other resolution
1024 x ??? or ??? x 1024
=========================================================
+I'm currently doing research on FLUX.1 lora. This is going to take a very long time.
SDXL(PONY, Not optimized) 24.08.19
+I'm going to improve this lora soon
-Trigger Prompt) LLF, LALAFELL *Use either or both. LLF is a very powerful trigger prompt. It's also a good way to not use this depending on the situation.
-[optional] Prompt) source_anime, LLFART
-Sampling method) Euler/Euler a
-[optional] Negative Prompt_pony hidden prompts) (gpo, aca, aer, api, fla, gcx, hll, hnj, gpc, fii, fey, fbv, evg, iew, ifl, igh, iwj, iwp, ixb, ixe, ixz, jaf, jbm, jfb, jsf, jyk, kmz, ksh, kxg, kzg, lbv:0.4) -> You can use this prompt even if you don't use this Lora. Users of the Pony model who help to improve the quality but were unaware of this may be disappointed by the changed image :)
I played FF14 at 6K resolution for collect the dataset
It greatly improved the quality of the hands, but the feet did not improve significantly due to the limitations of the dataset (Please add a foot prompt to address this to some extent (e.g sandals etc))
Still some prompts have problems with thin legs(e.g skirt lift, sitting). This is a problem caused by a lack of datasets (I hope this will be resolved in the next version)
This Lora was made in a very special way, and one of them is [LLFART] prompt
It's a model that's not fully optimized yet. This is a problem caused by the fairly good body representation of the pony model.(Lalafell has different body proportions to the another characters)
LLFART prompt (for PONY.ver)
- It was used to train for lala's body, But this increases the chances of creating bad hands. It should be used depending on the situation. Generally, it's used for negative prompts or not used anywhere.
ver.4(Beta/Beta_overfitting) 23.11.27
It's a difficult Lora to use. Detailed prompts are required (style, pose, etc)
There is a problem with the clarity of the face because i used a lot of low resolution images. It is recommended to use extensions (ADetailer) or to use Hires.fix together
Some poses or images may cause problems
Better results in various poses and prompts
Improves representation of body proportions
Delete Trigger Prompt
overfitting) Lora over-learned some elements of Lalafell. If you use this well, you can expect better performance
SDXL ver.C 23.09.14
Better results in various poses
Improves representation of body proportions
[optional]Resolution 1024 x ?? or ?? x 1024
[optional]Negative Prompt) toddler
There is a high probability of better results than version A and B. However, there may still be problems with some resolution and aspect ratio.
SDXL ver.B (pic lalafell) 23.09.12
Better quality at most resolutions
Better results in various poses(It's the result of sacrificing some of the characteristics of lalafell)
Prompts) lalafell, (optional)tall
Negative Prompt) toddler
(Low resolution) Use Hires.fix
Relatively high probability of body proportions compared to A version may result in mismatched body proportions (=unlike lalafell)
SDXL ver.A (ingame lalafell)
*Beta -> A.ver
Prompts) lalafell, (optional)tall
Negative Prompt) toddler
(Low resolution) Use Hires.fix
ver.3(SD 1.x)
You can apply more complex prompts + high resolution, and you can express the shape of the lalafell body more clearly (However, detailed prompts are required. It's hard to use)+Also, it may not be your taste because it is different from ver.2
Trigger prompt) lalafell
[optional] Prompt) female child, chibi (This makes lalafell look like a short lalafell in the image. If it seems too short, exclude chibi)
Negative Prompt) cat ears
ver.2(SD 1.x)
Unnecessary elements have been deleted (It shows better quality)
Trigger prompt) lalafell
ver.1(SD 1.x)
Lora made with selected images. (Images for learning only lalafell)
Trigger prompt) lalafell
Description
It is less likely that unnecessary elements will be included in the image.
FAQ
Comments (6)
sdxl as in its original form is the hugest pile of shit you never even should try to use.
use every other checkpoint but never sdxl itself that can be used as xl checkpoint.
Hah I agree with that. I'm currently waiting for a new version of WDXL.
I think we'll have to wait a little longer.
lalafell pony pls, thank you :3
If you're talking about the pony sdxl model, I'll give it a try. I was going to use animagine 3.0 or 3.1, but I realised that there are a lot of people using the pony xl model. I'm working on tagging for the pony model, so hopefully this will lead to good results.
Love-hate relationship with the [Beta] Pony lora. been using it for a while to establish the right body shape and face structure, but even the underfitting model is overtuned.
Art can be uncanny resembling ingame shots, mostly resolved by not using the Lalafell tag as you directed... But then you start to lose the body type.
Unprompted accessories and such appear and using your negatives excessively to resolve this leads to deep frying.
Styles and compatibility with other loras is pretty low, I'd say. though I dont know a lot about that aspect, or loras in general.
I think it's just a poor dataset but I haven't seen the images or what have you.
I'd love to contribute some better images of taters if your dataset is lacking and you're looking to re-train for the out-of-beta release. Until then I'll keep using this ver.
I agree with you. Currently the beta version has such a problem.
In fact, most of it is due to a few(or poor) data sets. I was planning to collect more than 500 QHD 4K images, but now I'm learning with about 100 images. So if you can provide a lot of data sets, I can improve it. hah
Before that, I'm going to make modifications so that the images I've collected can be used as datasets. Maybe I'll write this comment and upload the next test version if there are no big problems with the version I'm testing.