Close to my other lora : Ghibli style (Kiki's Delivery Service), but with a retro vibe, more grainy and less vibrant.
Krea2
Showcase images :
Turbo model
768x1152 - x1.5 Hires (10 steps, 0.2-0.35 with 4x_foolhardy_Remacri)
Euler / simple / 10 steps / CFG 1
Lora strength 1
Klein 9B
Showcase images :
Distilled model
768x1152
Euler / normal / 4 steps / CFG 1
Lora strength 1
Qwen
2512
It was trained on 2512, but I prefer to use the original Qwen-Image model for inference.
Showcase images :
Original Qwen-Image (BF16)
Lightning lora (4 steps)
5 steps / 1 CFG
DDIM & Beta
Lora strength 0.9
v1
All the showcase images have been generated without upscaling :
1584x1056 (portrait and landscape) with ddim / beta + Lightning lora
5 steps and CFG 1, and a lora weight of 1.
I've only generated 2 images for each prompt, and picked one, so it gives great results without cherry picking nor upscaling !
Flux
Trained on 29 carefully handpicked images from the movie Porco Rosso.
With a weight of 0.8, I start the prompt with :
ghibli style image of
All images are x1.5 upscaled with 4x_NMKD-Siax_200k (denoising strength 0.35 - 0.5).
Description
FAQ
Comments (5)
Any insight to how you trained this one?
It REALLY holds up with with Qwen Edit 2509 and I'd love to know why haha
Is that training caption correct? 4400 steps and 4 epochs? that would imply 38 repeats/batch size which seems wild
I've trained online on Tensor.art with these settings :
Qwen Image (BF16)
Dataset of 55 images (here one caption example : https://i.imgur.com/FvGRy6I.png)
LR of 0.0006 (20 repeats / 4 epochs / Cosine scheduler)
16 / 8 (Network Dim / Network Alpha)
4 / 1 (conv_dim / conv_alpha)
And I know the LR seems really high, but i've trained several qwen loras with these exact settings (style and character) and it was always good
@Yofaraway Awesome thanks I really appreciate the reply! Going to try these settings now on a stubborn set.
@Yofaraway Isn't she touching the piano with her right hand, not left? Or is that how training sees it since it's on the left side of the frame? Just curious. I have no idea how to do any of this. :)