Replies: 1 comment 13 replies
-
you named your images something like 0000-0-a 0.1.jpg, with the fast method, the instance name/prompt would become "--a" which is a bad instance name, it interferes with the letter "a". And you didn't summon my method in your prompt because you didn't use the instance name. The results you got from my method are not actually the result of the training. Try again the NEW method by renaming the instance pictures to "jjgkkgjgkl (1) .... etc" and in the A1111, call the method by using "jjgkkgjgkl" instead of "ba-shiroko". |
Beta Was this translation helpful? Give feedback.
13 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
I trained the following character images using Fast DB and ShivamShrirao DB and compared the results.
https://github.com/ShivamShrirao/diffusers/tree/main/examples/dreambooth
Both trained at 1500 steps and took similarly about 20 minutes.
The result is as follows.
Euler, 28steps, CFG 11, Clipskip 1
prompt : ba-shiroko, masterpiece, best quality
negative : nsfw, 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
The four on the left are Shiv, and the four on the right are Fast DB. Bikes and gloves included in the imageset appear stronger in Shiv.
The Regularization-Images I used with shiv are this.
Beta Was this translation helpful? Give feedback.
All reactions