Quote:
Originally Posted by drewprops
…but training the AI seems to be the key and it sounds like you need a blue million photo references?
|
If you wanted to train a whole model from scratch, then yeah, you need a millions pictures and a
many thousands of dollars worth of hardware-time. Very few folks are creating whole models from scratch. What I'm trying to do is create a "textual inversion" which is an add-on file that influences the behavior of an existing model. My TI output is a
tiny file that's only 4 KB compared to the multi-GB trained model files. From what I've read, you only need a dozen or so pictures to make a good TI, but like with everything else about these systems, the quality of its output is highly dependent on…
1. the quality of the pictures with some variety (different angles, backgrounds, clothes, facial expressions)
2. good textual descriptions associated with each picture
3. a
lot of patience
(and willingness to throw it all away and start over when it fails)
I was struggling to train/build the TI file on my own system. So, I used free credits on a Google Colab server to crunch the numbers. It took
three and a half hours to process twelve 512x512 pictures, and I had to keep poking things on the page every few minutes to make sure it didn't think I was "idle" and shut itself down because their free credits are not intended for long-running, unattended jobs like this. Yes. It took 3.5 hours to make a 4 kilobyte file. The maths are weird.
In hindsight, I think I framed my images too closely on his face, and I didn't provide enough detailed descriptions for them. If I do this again (and I probably will), I'll try to run it all locally so I can let it crunch overnight worry-free.
Quote:
Originally Posted by drewprops
EDIT: Worf of the Thousand Ridges  |
I think Chewbacca-Worf near the center might be my favorite unexpected reject.