Hi, I used rank 1 for this lora, that makes it a smaller size. Rank 1 around 10mb, rank 2 20mb...

@cynicalxrose Hi, no problem. I have learnt it from others in ai-toolkit discord server and some articles but I didn't save them, sorry.

I can try to explain it to you though but don't trust me 100%, I'm not expert!
The rank or dim/alpha would be like the coverage of a paint coat, a higher value will have more capacity to influence the original model, with higher rank the lora will have more parameters to capture more nuances.
To store the additional parameters the lora file will grow in size. In theory it would result in higher quality, with Flux the difference is not too noticeable for me but maybe I have a bad eye so take it with a grain of salt.

The layers or blocks, are the stages the model goes through in inference, Flux has over 50 of them, supposedly each one takes care of a different aspect of the image, but this info hasn't been made public by Flux creators. There are experiments with training in only a few blocks and it works for some things, since you are not training all parts of the model it also reduces the lora file size.

I hope this helps! I recommend you join the discord server, there a lot of knowledgeable people there https://discord.gg/VXmU2f5WEU