ComfyUI

Training a lora (stable diffusion) [part 1]

Step by step Lora Training

Training a model is very demanding, and the result is a minimum 2GB file ; while a Lora (Low-Rank Adaptation) will be less than 150MB.

Where to start?

I would recommend this google colab link

Linaqruf/kohya-trainer/blob/main/kohya-LoRA-dreambooth.ipynb 

Step 1:

Requirements: The preparation of a set of images, and texts (dataset)

Another requirements is a google account (to use the colab link and google drive)

Each image and text's name must match - the size of the images should be 512x512 pixels 

It is convenient to use https://www.birme.net/ to resize the dataset.

For the text, best is to give a little description on each file, or to use the img2txt feature and get the terms directly from the AI.

 Only 5 images (and texts) are going to suffice, but 15 or more will make the model more reliable.

Once the dataset is complete, put these files into a folder with the number of repeats (5 or 10) and the concept (something like exp, or person, or Bcloth, etc ...)

Once ready add the datasets into the google drive

Step 2: 

Open the colab trainer

Go to github.com/linaqruf/kohya-trainer

and pick Kohya Lora Dreambooth

To be continued ...

Creating content is so easy now

Cartoonists, musicians, voice actors, and more broadly every artists or content creators are now in competition with AI generated contents.

This is a game changer, and we have not completely understood or contemplated how much of a revolution we are seeing emerging ... in early 2023

It took me less than 10 minutes to create a new super hero from scratch, write a backstory, and several scenarios, in less than a day, I could self publish an entire comic book featuring this new Dark violet super hero ...

I used lexica.art, dreamlike and stable diffusion to get a different artstyle

AI and singularity

Grammar and syntax as a filter

Using a stream of knowledge base (be it erotic literature, postmodernism, or legal documents) through grammar and syntax filters would enable an artificial intelligence apply to generate an infinity of text combinations, it is in essence what the megahal project is doing, using the previous dialogs as a database for further conversations.

 Here is an example of automatic generated text (with the link below)

'Do you have anything in mind? ,' he whispers, his voice low, and dark desire unfolds throughout my body as everything in my body tightens.

'Okay,' I breathe .

'Good,' he growls, his breathing harsh.

'Don't make me wait,' I whisper.

He runs his nose along my jaw, and I'm quaking like a leaf. His mouth is on mine, tasting, exploring. He plants a soft wet kiss on my lips, and a pulse of warmth pools inside me. He plants a soft kiss on my lips, his eyes glowing with lust, raising uncontrollable urges in me. His hand tightens around my hair at my nape, pulling my head back as we kiss. My subconscious glares at me.

generator is : http://www.xwray.com/fiftyshades?id=2341

Another interesting generator for post-modernism article : http://www.elsewhere.org/pomo

But as always, the catch is that the AI doesn't have a clue of what make sense or not -- that's why singularity scenario where robots or AI are taking over the world are completely idiotic and unrealistic.
 
#artificial.intelligence
#automated.text
#limitation