The original base model may have already been trained with images in this particular style, but the tag associated with it doesn't perform well and it tends to produce many red metallic creatures. Therefore, here is a style LORA model that has been trained on a dataset of over 1000 images.
Description
32/16(dim/alpha)
adafactor(scheduler/optimizer)
1/0.1(scale weight norms/network dropout)
true(de-biased estimation loss)
1(batch size)
Caption data is combination of tags fully scraped from image board by Grabber and tags generated by wd-eva02-large-tagger-v3.
Model is initially trained for 9 epochs after roughly removing irrelevant tags via Tag GUI. Then, duplicated and style-error images are removed. Additionally, tags associated with numbers are fixed and many instances of "armored bodysuit" are added to the dataset. After this process, training resumes from a saved state for 8 more epochs. It is found to have strong style adherence.
FAQ
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.