Thanks!

> This model has 1 file scanned as unsafe. testvl-pre76-top187-rec69.pth

Hm, perhaps I'll wait for this to get cleared up?

This is not the first time I've heard of checkpoints being used to distribute malware. In fact, I've heard this was a popular vector from shady international groups.

I wouldn't expect this from Bilibili's Index Team, though, given how high profile they are. It's probably(?) a false positive. Though I wouldn't use it personally, just to be safe.

The safetensors format should be used by everyone. Raw pth files and pickle files should be shunned and abandoned by the industry. It's a bad format.

Disty of SD.Next has made a version in diffusers format.

https://huggingface.co/Disty0/Index-anisora-5B-diffusers

For the record, the dev branch of SD.Next (https://github.com/vladmandic/sdnext) already supports it.

thanks

I wonder if the entropy of model weights and their size causes statistical false positives to appear often?

I imagine it has more to do with whether or not the file appears to have executable python code in it, as a .pth file is usually just a a pickled python object and these can be manipulated to load arbitrary python code when loaded.

> Not sure it's wise to call it `AniSora` but I don't really know.

Given that OpenAI call themselves "Open", I think it's great and hilarious that we're reusing their names.

There was OpenSora from around this time last year:

https://github.com/hpcaitech/Open-Sora

And there are a lot of other products calling themselves "Sora" as well.

It's also interesting to note that OpenAI recently redirected sora.com, which used to be its own domain, to sora.chatgpt.com.

> OpenAI recently redirected sora.com, which used to be its own domain, to sora.chatgpt.com.

Probably to share cookies.

Cookies are such a mess.

We need cross-domain cookies. Google took them away so they could further entrench their analytics and ads platform. Abuse of monopoly power.

We use OAuth2 for identity.

We use first-party cookies for session management.

We use APIs and signed tokens (JWT) to federate across domains without leaking user data.

The ones hurt by the death of third-party cookies are ad tech parasites who refused to innovate imho...

OpenAI uses all of these things and still had to move domains.

Also: tech should be easier, not harder.

Building this shouldn't take more than an hour, yet somehow we did this to ourselves.