Jamie

Doing a quick skim on my phone, your microphone quality is fine. I would probably lower the game audio in post a bit to make the sound more distinct, but it's only noticeable when the game does loud stuff.

Speaking for LLMs, given that they operate on a next-token basis, there will be some statistical likelihood of spitting out original training data that can't be avoided. The normal counter-argument being that in theory, the odds of a particular piece of training data coming back out intact for more than a handful of words should be extremely low.

Of course, in this case, Google's researchers took advantage of the repeat discouragement mechanism to make that unlikelihood occur reliably, showing that there are indeed flaws to make it happen.

Accumulated knowledge in our society really is frail. Take a computer mouse, tons of people are involved in making them, they're considered extremely simple tools. Yet not one person on the planet could go out into nature, get the natural resources required, and without help turn those resources into a working computer mouse.

I'm not an expert, but I would say that it is going to be less likely for a diffusion model to spit out training data in a completely intact way. The way that LLMs versus diffusion models work are very different.

LLMs work by predicting the next statistically likely token, they take all of the previous text, then predict what the next token will be based on that. So, if you can trick it into a state where the next subsequent tokens are something verbatim from training data, then that's what you get.

Diffusion models work by taking a randomly generated latent, combining it with the CLIP interpretation of the user's prompt, then trying to turn the randomly generated information into a new latent which the VAE will then decode into something a human can see, because the latents the model is dealing with are meaningless numbers to humans.

In other words, there's a lot more randomness to deal with in a diffusion model. You could probably get a specific source image back if you specially crafted a latent and a prompt, which one guy did do by basically running img2img on a specific image that was in the training set and giving it a prompt to spit the same image out again. But that required having the original image in the first place, so it's not really a weakness in the same way this was for GPT.

I consider it occasionally, then remember I'm paying a ton more to save like, 15 minutes. Then I just go get it.

I'm not talking strictly about ideas, I'm talking about a human having a vision, and taking action to make that vision into something. Whether something is copyrightable requires a "human element," which is the reasoning behind why machine or animal generated content cannot be copyrighted, because they lack that.

So the question is if someone tweaking an image, even if they're merely selecting things, then is that a sufficient human element to say that a person had enough hand in creating it?

When it comes to selection, we already have a valid form of copyright which is explicitly that- compositions. If I take a bunch of royalty-free songs, and make a book of sheet music where I hand selected songs to be in that book, I can own a copyright on the composition without owning any of the featured material.

So, if someone selects a bunch of individual elements in an image using img2img, is that now a composition?

I accidentally submitted early, but also, I wrote out the lyrics. It's the most bland version of those breakup-depression kind of songs imaginable. I guess people voted it as "feel-good" out of irony.

Sitting at my favorite cafe

Sipping my tea it's saturday

Thinking about all he's done, to everyone

This town is full of broken dreams

Shattered hopes, and silent screams

Somebody please help me

Betrayed by this town

Let's tear it all down

We're all just destined to fall

I've lost it all

Betrayed by this town

Let's tear it all down

We're all just destined to fall

We've lost it all

Alone in the streets, alone in my thoughts

Thinking of all our favorite spots

I thought someday things might turn around

But I was lost and never found

Betrayed by this town

Let's tear it all down

We're all just destined to fall

I've lost it all

Betrayed by this town

Let's tear it all down

We're all just destined to fall

We've lost it all

Faces painted with smiles

Lies are told

A facade of unity

A vitality sold

So I sit here in silence

Just wondering how

To rewrite the tales

This town won't allow

Betrayed by this town

Let's tear it all down

We're all just destined to fall

I've lost it all

Betrayed by this town

Let's tear it all down

We're all just destined to fall

We've lost it all

I've lost it all

We've lost it all

I have a feeling they knew how this would be received considering it seems like they're rage-baiting and acting pretentious to try and get attention.

Some AI generated images can require a lot of tweaking to get a final result. For example, someone might have a workflow that involves generating several images, then picking one as a base. They then take that base, and use img2img to rework certain parts to suit a vision before applying a set of post-processing effects in a traditional editor.

Or, they generate an image and use it as a base for some sort of more traditional art, or use AI generated elements in a work that is otherwise drawn traditionally.

There's a lot of grey where I think just dismissing any creative vision is doing disrespect to the person that wanted to make something out of that vision, and put in a good amount of work outside just proompting and taking the first image that looked okay.

The issue, I think, was because most of what I use it for is anime. So some shows wanted the Japanese title, others wanted the English title, some couldn't be found at all. My US TV shows and movies never had that problem.

The title matching is what made me go to Plex. Some shows were impossible to get sorted right on Jellyfin. Plus there's a lot more ecosystem around Plex