I think it's important for people to diversify away from them and not build anything that uniquely depends on them. It's not good to have chokepoints like this.
I don't quite understand how they do so, but it seems to be possible. Everyone who downloads a 5GB stable diffusion model probably downloaded 5GB of pytorch+cuda+cudnn first.
This itself is kinda insane. In fact, it was literally untenable for PyPi, which is why (historically) installing all that was unecessarily painful.
[1] https://dustingram.com/articles/2021/04/14/powering-the-pyth...
python -m spacy download en-core-web-lg
etc
If they reach a point where they actively become community-hostile, someone will just fork their codebase and release a web app called "FaceHugger"
What they do isn't too hard to replicate, but the price for which they do it is impossible to compete with.
Rumor has it Github is actively working on something similar to HF Hub already
pick one if you are lucky. Those Fortune-12 giants can't do anything "in a heartbeat" .. listen to what that means.. nonsense
Someone utilize p2p for the love of god. Pretty sure there’s gigabytes of porn moving (freely) through the internet millions of times per day.
P2P model distribution isn't impossible, but it's a lot less interesting to many people, or they would've already spread their models through torrents in the first place.
But if HF's server code is open source it should be trivial to make one.
In fact, I think the whole community is HF averse. The two most popular frameworks are based around the Stability implemenetaion and file format, not HF diffusers.
Automatic1111 and ComfyUI implemented the SAI backend/format in the early days because thats all there was, and now they are stuck with it. The intertia is tremendous.
It seems like an odd move to me as what makes Auto1111 so appealing are the extensions and I’m assuming that breaks an awful lot of them, but what do I know?
Technology-wise, path dependency from stuff that was built out before HF diffusers were available, and on which more continues to be built in the ecosystem.
If you are doing a greenfield project that is mostly standalone, its not an issue, but for existing popular projects and the communities of satellite projects around them, the switching cost is high.
In terms of content hosting, well, a fair amount is hosted on HF, but there is a difference of content focus between CivitAI and HF, and a lot of what CivitAI hosts HF probably wouldn’t want to. Also, CivitAI has a UI focused on the narrow space of imagegen, whereas HF is more general.
On top of that HF is just hard to use for your average user. Civit.ai is just "click to download" while HF is "look here's a broken model card... you can figure it out from here".
Despite the cute logo, I think most people find that HF comes across as fairly anti-user. Despite having years doing ML related work, I still find HF a bit byzantine to navigate.
https://github.com/huggingface/safetensors https://huggingface.co/blog/safetensors-security-audit
I don't know much about it, and I think (but couldn't quickly confirm) it's open source, so your point about a fork still stands, although I don't think that solves everything. If it did, people wouldn't care that e.g. Hashicorp changed their license away from open source.
There are network effects in play here. They have a big moat purely because of their large user base.
Model hosting is a nice convenience, but if HF removed every single repo from GitHub tomorrow and paywalled the model repo, it wouldn't be a big deal. Maintainers would clone their models and repos to somewhere else.