v2

Enabled by prompts such as anime, manga, etc.

It seems that some thought needs to be given to the prompts in order to generate high quality images with Flux.

The prompts used in the images posted on this page were generated using LLM based on the contents of the following URL.

https://huggingface.co/datasets/k-mktr/improved-flux-prompts-photoreal-portrait

416 x 608, 30 images.

No caption.

sd-scripts : fine tuning

optimizer : adafactor

scheduler : constant

LR : 2e-6

batch size : 30

epochs : 400

repeats : 1

I initially trained on a datasets of 74 images as well, but failed due to anomalies in the appearance of the hands.

I then reduced the number of images to 30 and the quality improved.

The cause is unknown.

v1

Test Version

Enabled by prompts such as anime, manga, etc.