Prompt with character name + (blue archive):
As an example, the prompt for the following grid is
Positive: (masterpiece), [...], (beautiful detailed eyes)), highres, extremely detailed CG unity 8k wallpaper, breeze, Flying splashes, Flying petals, wind, sakura, solo
Negative: (realistic, photorealistic), lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, bad feet, big boob, big breasts, simple background, three feet, extra leg, missing arms, (poorly drawn hands), nude, nipples, multiple girls, three arms
Training details:
Trained with kanewallmann Dreambooth repository using tags as captions
The dataset contains 200-400 images for each character and there is also a customized reg set with a few thousands of images
- I first trained for 25000 steps at lr=1e-6. The result was not very good. (19 characters here)
- I guessed it was a problem of character name and prepend everything with BA and trained for another 20000 steps at lr=1e-6. The result was better but still not good for all the characters. I also found prepending BA might not be a good idea. (still 19 characters)
- I added more images on the characters that the model was not good at, prepended BA only with probability 0.3 and lowered the learning rate to 5e-7 and trained for 17500 more steps. The result came out ok.
- At this point I was afraid that the model was overtrained and as it seemed that not adding tags could help. I do a 0.7-0.3 mix with the base model, added three characters (hifumi, iroha, ui), and decided to drop tags with probability 0.3 (only charcter name and copyright remained). Another 24000 steps at learning rate 5e-7.
- I guess the model is still overtrained. Since the last three characters were only added in the last 24000 steps the model did not learn ui very well.
More example generations:
).png)
.png)
.png)
