We’re so close to finally being able to generate our own Haruhi season 3… what a time to be alive.
veonik · 1h ago
Dude… are you telling me it isnt actually finished? I am watching season 1 for the first time…
isaacimagine · 4h ago
I tested this out with a promotional illustration from Neon Genesis Evangelion. The model works quite well, but there are some temporal artifacts w.r.t. the animation of the hair as the head turns:
> a variable-length training approach is adopted, with training durations ranging from 2 to 8 seconds. This strategy enables our model to generate 720p
video clips with flexible lengths between 2 and 8 seconds.
I'd like to see it benched against FramePack which in my experience also handles 2d animation pretty well and doesn't suffer from the usual duration limitations of other models.
So was this trained on existing anime? Ain't no way the corpus was licensed legally.
tonyhart7 · 2h ago
"animated video generation model presented by Bilibili."
You understand that china has "different" view on copyright,license etc right??
yorwba · 8m ago
Not that different. Bilibili is a big, above-board video streaming service; they definitely have distribution rights to a large collection of anime content. (They also have YouTube-style user uploads where proper licensing is less likely.)
It's the equivalent of Crunchyroll putting out a video generation model. If the rightsholders disagree with this usage, it'll come up during the negotiations for new releases.
mythz · 1h ago
China doesn't know what you're talking about.
Lerc · 1h ago
The right to train models on copyrighted data has yet to be determined.
mattigames · 31m ago
Not like chatgtp and sora which as all we all known are fully trained in public licensed content free of copyright.
ekianjo · 40m ago
There are very few models out there that are not trained on data protected by copyright. So nothing new for the past 3 years
ronsor · 1h ago
China doesn't care about silly licenses.
MattRix · 23m ago
I might be missing something, but it feels weird that it’s named after Sora?
chii · 19m ago
sora is the japanese word for sky, and it's not that uncommon a name.
smusamashah · 3h ago
There are so many glitches even on the very first example. Arm of the shirt glitching, moving hair disappear and appear out of no where. Rest is just moving arm and clouds.
washadjeffmad · 3h ago
>Powered by the enhanced Wan2.1-14B foundation model for superior stability.
Wan2.1 is great. Does this mean anisora is also 16fps?
babuloseo · 3h ago
So we can finally remake Akame Ga kill?
throwaway314155 · 4h ago
Says it's open source but I'm having trouble finding a link to weights and/or code?
Looks incredibly impressive btw. Not sure it's wise to call it `AniSora` but I don't really know.
> This model has 1 file scanned as unsafe. testvl-pre76-top187-rec69.pth
Hm, perhaps I'll wait for this to get cleared up?
userbinator · 2h ago
I wonder if the entropy of model weights and their size causes statistical false positives to appear often?
throwaway314155 · 1h ago
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.
echelon · 4h ago
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.
echelon · 4h ago
> 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:
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.
pests · 19m ago
> OpenAI recently redirected sora.com, which used to be its own domain, to sora.chatgpt.com.
Probably to share cookies.
s0rr0wskill · 2h ago
can i generate hentai
topato · 1h ago
Inquisitive minds need to know!
But seriously, I had the same thought, considering the general lack of guardrails surrounding high-profile Chinese genAI models... Eventually, someone will know the answer... It's inevitable...
https://goto.isaac.sh/neon-anisora
Prompt: The giant head turns to face the two people sitting.
Oh, there is a docs page with more examples:
https://pwz4yo5eenw.feishu.cn/docx/XN9YdiOwCoqJuexLdCpcakSln...
> a variable-length training approach is adopted, with training durations ranging from 2 to 8 seconds. This strategy enables our model to generate 720p video clips with flexible lengths between 2 and 8 seconds.
I'd like to see it benched against FramePack which in my experience also handles 2d animation pretty well and doesn't suffer from the usual duration limitations of other models.
https://lllyasviel.github.io/frame_pack_gitpage
You understand that china has "different" view on copyright,license etc right??
It's the equivalent of Crunchyroll putting out a video generation model. If the rightsholders disagree with this usage, it'll come up during the negotiations for new releases.
Wan2.1 is great. Does this mean anisora is also 16fps?
Looks incredibly impressive btw. Not sure it's wise to call it `AniSora` but I don't really know.
> This model has 1 file scanned as unsafe. testvl-pre76-top187-rec69.pth
Hm, perhaps I'll wait for this to get cleared up?
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.
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.
Probably to share cookies.
But seriously, I had the same thought, considering the general lack of guardrails surrounding high-profile Chinese genAI models... Eventually, someone will know the answer... It's inevitable...