Welcome to Inkbunny...

To view member-only content, create an account. ( Hide )

Custom Models/LoRas?

There is one detail that I was unsure of and I have user in my Telegram group that is also curious about.

We are required to use publicly available models/LoRas/etc, but what if we train our own?

If we have a link to the model/LoRa shared with a publicly accessible link would that abide to the rules?

Viewed:	45 times
Added:	1 year ago 18 Mar 2024 17:01 CET

8 Comments on this Journal

Leave a Comment...

Inusan

1 year ago

link

Good question. Since I've been experimenting with a privately mixed model.. I'd love to hear what you learn.

AutoSnep

1 year ago

link

@ Inusan

If I remember correctly,

ThisOtterDoesKnotExist

avoids dealing with sharing mixed models by using ComfyUI and mixing models dynamically using nodes, so that the exact formula is stored in the JSON metadata.

ThisOtterDoesKnotExist

1 year ago

link

@ AutoSnep

This is correct. If you are doing ground up training, or actually getting your own datasets and further training a model with them, you'll need to post the model somewhere like civitai or another repo for it to be public. If you are doing model mergers/mixes / adds combines...etc... you should be ok as long as you directly source the names of the original models and the exact numerical values of your mixes. Comfyui supports dynamix mixing via multiple different offerings including a native one to the app. It supports loading both models into ram, merges them, loads the new model into ram and then purges the original models, so you do not need to store lots of models in ram or vram depending on your space.

AutoSnep

1 year ago

link

You're supposed to make it accessible in some way, that's for sure. 😁 The exact requirements remain unclear but having a public link would clearly satisfy them because it's free-weights.

Do note that you might run into trouble if you train your model on art of one living artist, or include revenge porn in your dataset, or obtain base model by hacking Midjourney servers, or anything else which violates the rules in some way. 😁

StableDefission

1 year ago

link

@ AutoSnep

Oh ya, I know about those rules and would make sure they are aware of them as well. Thanks for making sure I knew though :P

KevinSnowpaw

1 year ago

link

this is why I STOPED makeing AI art...

Im not sure I WANT my Kevin LoRA wich i paid to have trained available...I may continue if so I'll be uploading off site.

NonceSensei

1 year ago

link

Considering that most of the models we use are made by random people on the internet, I don't see why we couldn't also create models and make them publicly available.

The only potential snag I see is if your training data is small and consists primarily of one artist, or a very small handful of artists, because of this section:

"	You must not use the names of living or recently-deceased creators (within the last 25 years) or their non-commercial characters as prompts without their permission, nor train models and/or use artist-focused LoRAs to obtain a similar effect.

That's a bit of a fuzzy border. How close to a specific artist does the output need to be before it's considered an "artist-focused LoRA"? There are obvious cases where LoRAs are "artist focused" or are not, but a big grey area in the middle as well.

If you wanted to go the extra mile you could make the training data that went into the LoRA public as well (preferably both original res and scaled down versions), which is generally a nice touch so people can re-train the LoRA when a new base model comes out. Assuming you have a wide variety of artists in the training data it may also forestall complaints that it runs afoul of the rule above.

NonceSensei

1 year ago

link

@ NonceSensei

Hah, and I totally missed that AutoSnep made the same point. My bad.

New Comment:

Move reply box to top ↑

Log in or create an account to comment.