31–40 of 237 posts
Earlier quoted context omitted.
LDM-400M was already able to generate text (predecessor of Stable Diffusion), thanks to the fact that every token in the text encoder (trained from scratch) was available in the attention layer.
>thanks to the fact that every token in the text encoder (trained from scratch) was available in the attention layer. >ChatGPT explain this like I'm 5
Neither the source code nor the weights are open source... This is actually worse than Stability AI's previous offering, in that regard.
They are technically open source. It's just that the model license prohibits commercial use and the code license prohibits bypassing the filters. So it's kind of worse than closed source in a way because it's like a tease. With no API apparently. Theoretically large companies or rich people might be able to make a licensing agreement.
Your second sentence contradicts the first. Prohibiting commercial use and prohibiting modification are each in and of themselves mutually exclusive being being "technically open source" (let alone both at the same time).
Earlier quoted context omitted.
They are technically open source. It's just that the model license prohibits commercial use and the code license prohibits bypassing the filters. So it's kind of worse than closed source in a way because it's like a tease. With no API apparently. Theoretically large companies or rich people might be able to make a licensing agreement.
> model license prohibits commercial use I thought that at first, but I think it only prohibits commercial use that breaks regional copyright or privacy laws.
16GB VRAM minimum is a bit steep. Sadly excludes my 3080 which is annoying because I'd like something better than Stable Diffusion locally.
If you don't mind the power consumption I noticed that older nvidia P6000's (24GB) are pretty cheap on ebay! My 16GB P5000 is pretty handy for this stuff.
New restriction in their License suggests the software can't be modified. "2. All persons obtaining a copy or substantial portion of the Software, a modified version of the Software (or substantial portion thereof), or a derivative work based upon this Software (or substantial portion thereof) must not delete, remove, disable, diminish, or circumvent any inference filters or inference filter mechanisms in the Softwar…
Then by definition it isn't open source, violating points 3, 4, and 6 of the open source definition. https://opensource.org/osd/
For anyone who doesn't know, DeepFloyd is a StableDiffusion style image model that more or less replaced CLIP with a full LLM (11b params). The result is that it is much better at responding to more complex prompts. In theory, it is also smarter at learning from its training data.
Not really, it's a cascaded diffusion model conditioned on the T5 encoder, there is nothing really in common, unless you mean that using a diffusion model is "SD style".
New restriction in their License suggests the software can't be modified. "2. All persons obtaining a copy or substantial portion of the Software, a modified version of the Software (or substantial portion thereof), or a derivative work based upon this Software (or substantial portion thereof) must not delete, remove, disable, diminish, or circumvent any inference filters or inference filter mechanisms in the Softwar…
I'm starting to get irritated by all these 'non-commercial' licensed models, though, because there is no such thing as a non-commercial license. In copyright law, merely having the work in question is considered a commercial benefit. So you need to specify every single act you think is 'non-commercial', and users of the license have to read and understand that. Even Creative Commons' NC clause only specifies one; they say that filesharing is not commercial. So it's just a fancy covenant not to sue BitTorrent users.
And then there's LLaMA, whose model weights were only ever shared privately with other researchers. Everyone using LLaMA publicly is likely pirating it. Actual weights-available or Free models already exist, such as BLOOM, Dolly, StableLM[0], Pythia, GPT-J, GPT-NeoX, and CerebrasGPT.
[0] Untuned only; the instruction-tuned models are frustratingly CC-BY-NC-SA because apparently nobody made an open dataset for instruction tuning.
[1] Insamuch as an AI model trained on copyrighted data can even be considered Free.
Neither the source code nor the weights are open source... This is actually worse than Stability AI's previous offering, in that regard.