Fashion design generation - Fine tuning question
Hi, I'm new to stable diffusion and wondering if it's possible to fine tune a base model like dreamlike photoreal 2.0 to match a specific clothing design, like a summer collection of a brand.
Also I'm trying to make the images come with a white background and in the same pose/style or something close to it (e.g a full body shot of a person standing in a regular pose).
Other thing I'm thinking about is to generate clothing images without a person in the photo, I was not able to achieve this just specifying it in the prompt, any ideas about how can I achieve it?
What I already have tried (and current trying):
LoRA with about 50 images of a person in a white background, same pose, full body shot. The idea is to have images in the same style but be able to generate new/good clothing for fashion insights.
Fine tune the whole model with ~32k images of fashion clothing (person included). The goal was to improve clothing generation. It took about 19 hours to train with a RTX 3090 24GB VRAM, and the results were not that good, actually the base model (dreamlike photoreal 2.0) were better.
The same as above with dreambooth, again the results were not that good. I still prefer the base model I'm using.
I would love to have some tips from you guys, the models I see here in CIVITAI are amazing and I'm wondering if I can achieve something similar.
Thank you in advance for reading!
In this situation I would focus my efforts on creating the best LoRA possible for that kind of clothing.
> but be able to generate new/good clothing for fashion insights
I would still train the original fashion style into a LoRA, and prompt to create the new content.
In the prompt you could attempt to generate novel styles by reducing the LoRA weight, or adding other tokens. If you add multiple fashion styles, what happens?
So if your LoRA is like
a person in a telobo dress <lora:telobo_dress:1> you could then try prompting
a person in a telobo dress with haute coture gown styling <lora:telobo_dress:1> <lora:Haute_Couture_Gown:1> and add other intentionally conflicting styles to the prompt