Klein 9B
Showcase images :
Distilled model
768x1152
Euler / normal / 4 steps / CFG 1
Lora strength of 1
ZIB
v1 :
Still experimenting with Z-image Base training. Showcase images :
Euler / normal / 20 steps / CFG 5
Hires steps : 20 / NMKD_SIAX @0.2-0.5 / x1.5-x2
768x1152
Lora strength 0.7 -1
Qwen
Qwen2512 v1:
First lora trained on the new Qwen2512 (it still working with previous Qwen-Image and Qwen-Edit).
For inference, I still prefer the first Qwen-Image model (BF16) (all the showcase images are done with this one).
Seems pretty good but few things appears sometimes unprompted in my tests: artefacts on skin (like scars) and japanese onomatopoeia
Qwen v1:
I've trained a lot of different versions for this one : with big and small dataset, trigger words or not, high and low LR... I've still no idea what's good or not. 🙂
Flux
Had good results starting prompt with :
'black and white Berserk manga style image of'
and 0.8 strength most of the time
v2 edit : Better dataset for better prompt adherence and flexibility
Description
FAQ
Comments (5)
Very interesting how close it looks to real manga art in some images. It still feels like Ai in most of them but it creates the best Berserk replica out of any ai tool. Nice wark, I wander if a bigger dataset could create better results. Oo how many images did you train on?
I often train on 20 to 30 images because it seems to give the best results and be the more flexible.
For another Lora I've tried 100 images, then 800+ images, and it was still the 30 images the best by far.
Flux is a strange beast when it comes to training compared to say SDXL. I found that in general, a small but high quality and consistent set produces the best result, because then the A.I. knows exactly what it is being taught, and it does not get "confused".
For something as specific as Berserk, maybe one can get better result by training a LoRA for each of the major characters?