The Voynich Ninja

Ok, I am now using Marco's prompt from the opening post. However, I can't test it yet because the GPU is still locked

(16-05-2022, 02:00 PM)bi3mw Wrote: You are not allowed to view links. Register or Login to view.What parameters do you have set ?

Hi Matthias, I attach my settings. As Keon said, I only edited the prompts (and the size).

Koen's images show that the software was mostly trained on photographs of 3d objects, not flat 2d images like manuscript pages. The results often tend to show some areas that appear to be higher than the level of the page, as maybe embroidery or some kind of thick painting.

While I find this software fascinating, I doubt it can have a direct relevance for the VMS. At most, it is one more proof of the amazing things that AI can do. I guess that people sufficiently skilled in linguistic AI could significantly contribute to voynichology (and they probably will, in the next few years).

Yeah, this particular application won't teach us much about the VM. But it is fascinating to have a look under the hood of new AI applications and see how they develop.

Marco said:

While I find this software fascinating, I doubt it can have a direct relevance for the VMS. At most, it is one more proof of the amazing things that AI can do. I guess that people sufficiently skilled in linguistic AI could significantly contribute to voynichology (and they probably will, in the next few years).


Well, darn -- but I knew I was being too naive to think we could just feed stuff in and have something helpful come out.  At least at this point.

That being said, I can contribute one interesting thing about this area.  It is actually a very hot topic as to who exactly owns the copyright of the produced images.

The reason this is an issue is that at least under US law, the owner is the "author" and creation is when the work becomes "fixed in a tangible medium."  Notably, the work has to be "original" but that is a pretty low bar.  Certainly, if it was a human was putting these images together, that human would be the author and the copyright in the created images would be owned by them. 

Is the AI program the "author" of the pages we see in this string?  There is already caselaw that a computer program cannot be an "author." But if not the computer program, then who?  The human coder(s) of the program?  The person who enters the parameters?  A combination?  

Certainly, Marco or Koen or Matthias "helped" with the prompts but is pretty clear they did not "fix it in a tangible medium" - that was the program.

I'm sure a few of you will think that there should be no copyright at all -- but that isn't really realistic given the system as it has evolved.  What if Marco and Koen and Matthias just wanted attribution credit?  There should be some sort of method for them to be able to require that which normally would involve copyright ownership.  But at this point, it is far from clear whether they own any copyright at all.  

It will be interesting to see how this all plays out . . .

As I understand it, in most jurisdictions the AI is simply deemed to be a "tool" and so the human who used the tool remains the copyright holder.
If no human can be assigned, then it falls within the public domain.
The UK's CDPA covers this, saying that computer generated work falls within the remit of the person who arranged the work.
It would be interested to see if the Terms of Use of such a programme could assign copyright....

The way it feels to me is that someone built a very nice piano and I played some notes on it. (i.e. my limited input). I guess the most sensible thing copyright-wise would be that I caused the thing to bring forth this particular result, so I would be the author of the image, even though I did not control its eventual appearance.

Here is a Youtube tutorial for the old version of Disco Diffusion, but nothing has changed in terms of operation.

You are not allowed to view links. Register or Login to view.

It should be noted that I used an initial image that I had to copy "by hand" via Google Drive to the following path:

/content/drive/MyDrive/AI/Disco_Diffusion/init_images/f77v.png

This is then also the path that has to be specified in Disco Diffusion under "Init Settings" / "init_image:".

Under "skip_steps:" you have to enter "125" if you have left the value of "250" unchanged under "Basic Settings" / "steps:".

In addition, the parameter "n_batches" must be set to "20" or less so that the entire process does not take too long and Colab does not, as in my case, lock the GPU for a certain time.

(16-05-2022, 05:23 PM)MichelleL11 Wrote: You are not allowed to view links. Register or Login to view.Is the AI program the "author" of the pages we see in this string?  There is already caselaw that a computer program cannot be an "author." But if not the computer program, then who?  The human coder(s) of the program?  The person who enters the parameters?  A combination?  

I think the situation here is analogous to compiler/source code/binary file problem. The author of the compiler is obviously not the author of the binary code produced (although thanks to tricky licensing they may hold the copyright?).

In our case AI engine is the compiler, and the prompt is the source code. The short length of the prompt by MarcoP (["A page from the Voynich manuscript", "ancient manuscript"]) is somewhat misleading, but even there he used "a page form Voynich manuscript", not merely "the Voynich manuscript". When the technology advances, the role of prompt and other parameters will be more distinct.

I wanted to try experimenting with an initial image, but that part didn't work for me. I tried putting a png in the same map you did, and then entered 'filename.png' in the initial image field, but this resulted in an error.

As far as I understand, the skip steps only have to be adjusted if you want to. Basically the closer it is to the total steps, the more of your initial image will remain.

(16-05-2022, 08:44 PM)Koen G Wrote: You are not allowed to view links. Register or Login to view.I wanted to try experimenting with an initial image, but that part didn't work for me. I tried putting a png in the same map you did, and then entered 'filename.png' in the initial image field, but this resulted in an error.

You probably need to enter the whole path. See Matthias' settings You are not allowed to view links. Register or Login to view..