DB training with captions?

[DJTcode]

For Dreambooth training with this repo, can I label the images with full length captions instead of a name?

So instead of titles like:

'JJohn (1).png'

I would use:

'a man wearing a blue jumper, JJohn, in a forest on a sunny day, photorealist.png'?

[TheLastBen]

Yes you can, separate the words with an underscore _ if you do the captioning, set the text encoder to 0%

[DJTcode]

Thanks

[SorenTruelsen]

Yes you can, separate the words with an underscore _ if you do the captioning, set the text encoder to 0%

BRO, i was on my 15000 step and google close the Ssession EVEN ON PRO. and NOW I CANNOT LOAD becuase of ERRORS IN CODE??? JESUS WTF?

Three things:

1. Your comment is not related to the an open issue and not remotely related to this issue
2. You should never need to train 15000 steps
3. You do not pay for this software, you have no right to complain.

[DJTcode]

2. ou should never need to train 15000 steps

Yes, the previous comment is not relevant to this thread, please vent your frustration elsewhere. However since its been mentioned:

Is there not a scenario where if you were to train a style or multiple concepts in Dreambooth you could train it beyond 15000 steps (if following a general rule of say 200 steps for every image with a low learning rate)?

[SorenTruelsen]

2. ou should never need to train 15000 steps

Yes, the previous comment is not relevant to this thread, please vent your frustration elsewhere. However since its been mentioned:

Is there not a scenario where if you were to train a style or multiple concepts in Dreambooth you could train it beyond 15000 steps (if following a general rule of say 200 steps for every image with a low learning rate)?

There is, true, training mulitple concepts or people at the same time requires more steps.

Training two people I found it much better to use 2x 10 images, for a total of 2000 than 2x 40 images for a total of 2000

[LIQUIDMIND111]

2. ou should never need to train 15000 steps

Yes, the previous comment is not relevant to this thread, please vent your frustration elsewhere. However since its been mentioned: Is there not a scenario where if you were to train a style or multiple concepts in Dreambooth you could train it beyond 15000 steps (if following a general rule of say 200 steps for every image with a low learning rate)?

There is, true, training mulitple concepts or people at the same time requires more steps.

Training two people I found it much better to use 2x 10 images, for a total of 2000 than 2x 40 images for a total of 2000

2. ou should never need to train 15000 steps

FYI i DID had an open case, NOW CLOSED, so chill.

PLus, have you ever done DEEP FAKES IN YOUR LIFE?? Do you have training experience outside dream booth??

i don't get the issue, its just MATH - "RECOMMENDED IMAGES 1 x 100 (or 200 actually) steps = 10 images then 1000 steps and 100 images then 10000 steps,,,, simple math, and YES if you use style and people combined then you need to work also the text encoder and steps count..... indeed. EVEN WORST, if you use a BIGGER MODEL like 768 then EVEN MORE TIME OR STEPS ARE SUGGESTED...

So Thanks for criticizing my training method then making an issue about it, cheers.... btw, i DID TIP BEN FOR THIS REPO... so CHILL. no need to me rude with your comments......