Does the SD3 license not allow fine-tuning the model without paying?
community not allow to finetune sd3
Not at all, I see this guys is update fine-tune it
repo: https://github.com/AMEERAZAM08/diffusers/blob/sd3/examples/dreambooth/train_dreambooth_lora_sd3.py#L1432
😂😂😂😂😂😂😂😂
You can finetune all you want. But you cannot sell the finetune or make money off the images produced by a finetune.
While it is basically impossible to prove that images were produced with SD3, it is fair.
Use it, tweak it, have fun...
If you want to make a profit, pay your share.
It's that easy
Hypothetically, what's stopping someone from fine-tuning and then claiming the model was trained from ground up, thus bypassing the license? Or is the entire architecture licensed? Hypothetical scenario of course.
SD3 uses several other open source projects as inspiration, such as Wuerstchen which is licensed as MIT: https://github.com/dome272/wuerstchen. so several parts of the architecture can be reproduced without requiring the license. if you have the compute power you can train from scratch (like RunwayML did for SD1.5). If you are interested in the mechanics, you can read the research paper: https://arxiv.org/pdf/2403.03206
Claiming other's work as your own is mostly an ethical issue, but let's be honest. A vast majority of people who train Models don't really care about ethics, so then there is still the legality of License abuse.
I am pretty sure that a diff of two models can clearly show if a model is trained from scratch or based on a licensed model.
Feel free to risk it and face the consequences.
If you make money on a finetune then paying a monthly 20$ should not be a problem. If it is, then your business model does not add up.
SD3 is running on MIT, but MIT isn't meant to be forcing people into a subscription to use it commercially.
This makes 0 sense