santa hat
deerdeer nosedeer glow
Sign In

Clip or unclip text encode? And more ....

Clip or unclip text encode? And more ....

Training settings:

config - bf16 archer.json you find as attachments.

Concept:

Samples from training:

8 epoch

and from 8 to 50 epoch

All data set (images, captions) atsached to article.

This is a imge without lora. Only prompt:

Now, connected lora, but not with text encode:

and got this:

Finally all connected:

and with simple prompt (still the same seed)

aderekarcher, woman, forest, spring

so, set lower str: 0.7

0.4:

without trigger (bow disapear):

and 0.1

So, lora is strong (overcocked) and work without trigger word "aderekarcher", lets check with low str and random seed: 0.4

0.7

0.9

and 0.9 with tag "aderekarcher".....

Come back to 0.7 and long prompt "aderekarcher, a woman in a hooded coat, holding a bow and arrow

arrow (projectile), 1girl, weapon, solo, bow (weapon), hood, gloves, realistic, brown hair, fur trim, snow, nose, lips, brown eyes, winter clothes, holding arrow, coat, snowing, blurry, holding, black gloves, quiver, upper body, winter, holding bow (weapon), holding weapon"

Im really not understood what happend here :)

so...

meybe next time will be better :)

And it is ;)

I was found "small" error in my testing and normal workflow.

You will find it in attachment section.

Have a nice day!

3

Comments