Didn't have enough time on this week) And Flux is bad at NSFW, need to wait NSFW Flux based checkpoints.
Last edited:
Generated with Flux 1. Dev, ComfyUI, without LoRA.
Code:
A photo of woman with purple hair, red eyes, short folded ponytail.
She holding wears armor with single pauldron.PC Specs: 32Gb RAM, 16 Gb VRAM, RTX 4080 Super.
Flux 1 Dev, ComfyUI, 1 additional LoRA
You don't have permission to view the spoiler content.
Log in or register now.
You don't have permission to view the spoiler content.
Log in or register now.
You don't have permission to view the spoiler content.
Log in or register now.
Flux 1 Dev, ComfyUI, 2 additional LoRAs.
You don't have permission to view the spoiler content.
Log in or register now.
To summarize..
Pros:
+tested LoRAs works;
+eyes really good;
+fingers decent (better than PDXL);
+size of LoRA file (variates from 9Mb to 350Mb or maybe more);
+can handle text;
+awesome details;
+supports ControlNet;
Cons:
-needs a lot of memory (RAM and VRAM), but there is a hope called NF4 compression (
-most of times checkpoint just ignores requests to add text);
-with LoRAs generation time can sometimes (moderately rare) be more than 30s (my worst one is 13 minutes);
-like to generate anime, when you ask photo or realistic image);
-bad at NSFW (LoRAs doesn't help much, because of their small sizes, need to wait specific checpoints);
-supports only positive prompt (negative prompt should be added later);
-generation time slows down significantly if you want text!
Pros:
+tested LoRAs works;
+eyes really good;
+fingers decent (better than PDXL);
+size of LoRA file (variates from 9Mb to 350Mb or maybe more);
+can handle text;
+awesome details;
+supports ControlNet;
Cons:
-needs a lot of memory (RAM and VRAM), but there is a hope called NF4 compression (
You must be registered to see the links
-most of times checkpoint just ignores requests to add text);
-with LoRAs generation time can sometimes (moderately rare) be more than 30s (my worst one is 13 minutes);
-like to generate anime, when you ask photo or realistic image);
-bad at NSFW (LoRAs doesn't help much, because of their small sizes, need to wait specific checpoints);
-supports only positive prompt (negative prompt should be added later);
-generation time slows down significantly if you want text!
Last edited:
It's difficult to get exact image, so i generated similar.
You don't have permission to view the spoiler content.
Log in or register now.
You don't have permission to view the spoiler content.
Log in or register now.
HailoKnight's model good only for close-up, for full body model generates different style, not photorealistic, last image looks more li drawing, the render.
I didn't ask to generate wings, but why not)
I didn't ask to generate wings, but why not)
Last edited:
Flux works better with backgrounds, a lot better!
But hands (especially fingers) still suffer)
Actually now i realize, that Flux if awesome, and has a lot of potential.
There is no chances that SDXL of PDXL can produce image with so detailed background. But still bad and deformed if compared to non-AI art)
But hands (especially fingers) still suffer)
Actually now i realize, that Flux if awesome, and has a lot of potential.
There is no chances that SDXL of PDXL can produce image with so detailed background. But still bad and deformed if compared to non-AI art)
Last edited:
- Jun 9, 2019
- 316
- 1,232
Next Pony would be based on AuraFlow instead of Flux(
AstraliteHearts (PDXL6 creator) comment:
Honestly it's amazing to even see Pony being in the same list as other heavy hitters but no Flux version in near future, I am focusing on AuraFlow version (more on this soon) given it's license, but I'll reach out to BFL and see how it goes.
You must be registered to see the links
AstraliteHearts (PDXL6 creator) comment:
Honestly it's amazing to even see Pony being in the same list as other heavy hitters but no Flux version in near future, I am focusing on AuraFlow version (more on this soon) given it's license, but I'll reach out to BFL and see how it goes.
You don't have permission to view the spoiler content.
Log in or register now.
You don't have permission to view the spoiler content.
Log in or register now.
You don't have permission to view the spoiler content.
Log in or register now.
BUT! Catch the goodies, gentlemen
-
I already lost faith in entire adult game industry.) And also not adult too.
thanks ill have a play around with it and see what i can make also thanks for the bigger image for my background
Trained 2 more KOD models, but this time with NN characters in data set. 1 training is fail, because i forget to set "num of repeats", so goodbye 1.5k buzz) Next one is better, but i tested only last (15-th) epoch. Riven looks like Sabia, because of red eyes, short hairs and folded ponytail.
All characters in training data set:
Females: sabia, neve, vehlis, ylva, kira, bris, maply, elmy, jasmine, amelia, lynn, riven, irelia, katarina, leblanc.
Males: dajrab, groknak, jadk, lutvrog, orcshaman, rokgrid, tekrok, avion, hessian.
Males all look gayish, because of bias to "bara" style in PDXL6.
You must be registered to see the links
Females +- decent, but they all mixed together, which means that they tends ti bias to each other) Ylva suffered more because of other orcs.
Because training for last one model was costed 8k buzz (because of 2048 training resolution), i don't want to further work with big models) So i think this is the last big model.
Model is stupid(. I will publish last epoch, just to add riven, irelia and katarina, but don't use this model to generate complex scenes with high strength.
You don't have permission to view the spoiler content.
Log in or register now.
Last edited:
I also publish last epoch from first training, because it slightly better, while less trained.
You don't have permission to view the spoiler content.
Log in or register now.
- Jun 9, 2019
- 316
- 1,232
Is your 3.5 version better than the 3.0 that I'm using?
You don't have permission to view the spoiler content.
Log in or register now.
Should be, for v3.5 i added high resolution images from packs, and crops with faces, also in high resolution. On civitai i mostly use v3ch because i want to keep clean gallery for v3.5) In A1111 i'm trying to use both versions.