Hacker News Reader

Comment by samtheprogram | Hacker News Reader

If I’m understanding this correctly, you can’t run this in a commercial setting, even if you’re not creating a derivative but simply generating outputs?

View on HN

I believe you can buy licensing? But def not the same as 'Open Weights..'

"Weights available" perhaps

The same people that claim using all of humanities creation is fair use want you to pay for a bunch of MatMul inputs that are unrecognizable to anyone after quantizing them yourself.

Stupid question, whats to stop someone from quantizing it, shit even just barely finetuning it for 1 step and calling it something different, no ones actually checking WTF these models are based on when they're released, especially for the source models, especially if the release is not around the same time of release as the base, i'm 99% sure someone could fine tune SD3.5 a bit and release it today as Frizz 1.0 and people would just take it as a new model using the same layer structure as SD3.5 lol

Not impossible but you'd gonna have to do a bit more than that. Most people are ignorant, but not all of them. An experienced user can tell what model family was used from a bunch of generated images. Also, no one would believe a nobody who just showed up claiming to have trained a brand new diffusion model.

I forget which, but some HiDream maybe was called out for this when it happened to generate basically the same dude in front of the same archway when compared against flux.

HiDream is a separate architecture. OTOH, it might be finetuned on FLUX generated data, we will never know.

HiDream is trained on AI generated outputs.

Yea, flux ones.

There is a simple method to detect this: taking a model "claimed" to be trained scratch, taking the model you suspected is the original, generate a new model = claimed_model * 0.5 + suspected_model * 0.5.

If the claimed_model is trained from scratch, the new model will have 0 capability (basically generate gibberish words or noise). If it is a derivative of the suspected model, it will do something sensible.

It is a bit more interesting for diffusion model because you can fine-tune to a different objective, making this investigation harder to do, but not impossible.

FLUX watermarks its outputs.

Additionally, certain prompts will produce nonsensical but specific outputs known only to BFL.

There is no watermarking in flux. The only artifacts that remain are vae artifacts. The vae is Apache licensed and used by many models now. So you can't identify the specific model.

I guess you’re in for a surprise if you make a giant application that uses FLUX dev!

lmao

The double standard is frankly disgusting.

I'm actually all for open training but I think it's only fair you treat the model as your treated the life's work of others.

Quite frankly, I still believe that these model licenses are dubiously enforceable at best, and I'm skeptical that models are copyrightable at all.

Reply on news.ycombinator.com