Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[MesoTroniK]

So far as I'm aware, this isn't terribly practical or likely to come up, because even building a basic LoRA that might produce "somewhat like" work takes quite a few images to even get within spitting distance. I think most of the people here don't know how hard that actually is. For the record, I've been completely uninterested in trying that; doing Dreambooth training just sounds like work.

But I'm with Alex; if it's done, and it's clearly infringing, they should be gone. I think we're all agreed on this.

This has indeed happened already, I have seen them.

[Wispborne]

In the same vein, is training a model (eg a SD checkpoint) on other modders' work without permission and releasing the output allowed? This has happened a few times already.

That's a good question. I'm tempted to say "no", though that runs into similar issues with how one might enforce this. So, I'm not sure. (I'm also tempted to say "yeah, you asked for this, so have fun!" But I won't, ahem.)
----(para removed to take less space)
Going a step further, I think training on the work of a specific actual artist to mimic their work, without their permission, is pretty poor behavior so if that comes up... well, it seems like just about everything with this would need to be on a case-by-case basis. But I could certainly see uninviting someone from the forum based on that alone.

Thanks! I agree with that stance. Training off of this community's work is a much more direct "harm" than what these models did, and in both cases where it happened, the people didn't disclose what they trained on until challenged. But if nobody can tell that someone trained on their work, even the original artist...it's a brave new world.

In the same vein, is training a model (eg a SD checkpoint) on other modders' work without permission and releasing the output allowed? This has happened a few times already.

So far as I'm aware, this isn't terribly practical or likely to come up

Thanks for your input, but it has come up already multiple times.

[BaBosa]

I guess the next logical progression of that question would be in regards to re-use of AI generated assets?  While it is hard for AI to mimic some particular person's style without feeding it a lot of samples, if AI generated art becomes the norm rather than the exception, it is very trivial for the same AI to generate more art similar to that originally generated if you've got the prompt and seed.  Reuse of similar AI generated images presumably wouldn't be a problem?  Likely not a problem for sprite ships, but it might be an issue for backgrounds and portraits.

I don't think someone can claim copyright over prompts and seeds so that's likely open game. If they worked over the images by hand enough that the AI can't remake it then that'd be there's in my mind.

[Hiruma Kai]

I guess the next logical progression of that question would be in regards to re-use of AI generated assets?  While it is hard for AI to mimic some particular person's style without feeding it a lot of samples, if AI generated art becomes the norm rather than the exception, it is very trivial for the same AI to generate more art similar to that originally generated if you've got the prompt and seed.  Reuse of similar AI generated images presumably wouldn't be a problem?  Likely not a problem for sprite ships, but it might be an issue for backgrounds and portraits.

I don't think someone can claim copyright over prompts and seeds so that's likely open game. If they worked over the images by hand enough that the AI can't remake it then that'd be there's in my mind.

So working from the same AI generated base and reworking it yourself, or using the original base would fine from that point of view though?

[Alex]

I *think* some courts have specifically said that AI output can't be copyrighted? So whatever came out of someone's prompt is not "theirs" in any real sense. I'm not 100% on this but that's my understanding. If that's correct, then reuse would be totally fine.

Thanks! I agree with that stance. Training off of this community's work is a much more direct "harm" than what these models did, and in both cases where it happened, the people didn't disclose what they trained on until challenged. But if nobody can tell that someone trained on their work, even the original artist...it's a brave new world.

*thumbs up* We'll cross some of those bridges when we get there, I suppose.

[xenoargh]

I guess the next logical progression of that question would be in regards to re-use of AI generated assets?  While it is hard for AI to mimic some particular person's style without feeding it a lot of samples, if AI generated art becomes the norm rather than the exception, it is very trivial for the same AI to generate more art similar to that originally generated if you've got the prompt and seed.  Reuse of similar AI generated images presumably wouldn't be a problem?  Likely not a problem for sprite ships, but it might be an issue for backgrounds and portraits.

Unless people are required to give the full prompts, seeds, number generators, etc. (which, frankly, is a bit burdensome and probably won't happen routinely) it's more likely that people will feed in existing images as img2img prompts and try to get "like that", I'd think?

In which case, it's back to, "is it clearly infringing?". I'd think that, for things like portraits and backgrounds, the results will be different-enough that it's a non-issue, unless it's kept so restrained re: prompt strength (essentially, how much noise is injected at the start) that it's 95% the same or better (and it'll be obviously a copy at that point).

This has indeed happened already, I have seen them.

Really? Who and how? Is this some Discord stuff that happened off-Forum? Or did I just miss it? I'm curious now, lol. If you just don't want to talk about it because it happened off-Forum, fine, but I didn't think anybody had the equipment, time or boredom required.

I *think* some courts have specifically said that AI output can't be copyrighted? So whatever came out of someone's prompt is not "theirs" in any real sense. I'm not 100% on this but that's my understanding. If that's correct, then reuse would be totally fine.

Actually, it's more complicated. The Copyright Office thinks that stuff like img2img works are probably copyrightable, but text-prompts aren't, unless they're incorporated into a larger creative work. Threshold appears to be "amount artist was involved directly with the creation". Fuzzy.

It's something that'll get tested in court a few dozen times over the next few years, to be sure.

[MesoTroniK]

This has indeed happened already, I have seen them.

Really? Who and how? Is this some Discord stuff that happened off-Forum? Or did I just miss it? I'm curious now, lol. If you just don't want to talk about it because it happened off-Forum, fine, but I didn't think anybody had the equipment, time or boredom required.

Yea is not forum stuff and I would prefer to not stir up that drama. But equipment? It isn't like 32gb of ram and a high end video cards are a rare commodity. Time and boredom? Haha, people are strange my friend.

[Void Ganymede]

There's a couple of experimental Starsector portrait LoRAs on civitai.com. They're objectively higher-quality than the anime/furry/etc portrait packs, but you can tell it takes a ton of cherrypicking to get something lore-friendly. Only the earliest seems on-model by default, rest stray quickly into other styles.

matrix

[close]

With enough cherry picking and creative prompting, some of the output's kinda neat. I bet this space witch has some stories.


Honestly hardest part of playing with LoRA is how janky and broken all the tools are. Given how fast the cutting edge is advancing it'll take time for them to stabilize.

[xenoargh]

OIC. You could actually do a LoRA that worked with that many samples. Barely, but.

Fair enough, lol. I'm really surprised anybody bothered, tbh. 32GB of RAM on a GPU isn't exactly common yet, outside really expensive stuff, so... you're saying somebody did that on a CPU? That must have taken foooooorever, lol.

You'd think with all the cool things you can get the tools to do, and the unlikelihood of getting decent results with low samples, why bother. I guess some people just had to find out, lol. Anyhow, fine, I don't need to hear about it and I agree that's not cool.

[Void Ganymede]

You can fine-tune with <12GiB VRAM. Tech is advancing faster than Python can be refactored or documentation can be written.

[xenoargh]

True. Every time I fire up Easy Diffusion, it has new features. This week they added multi-LoRA. Totally fun.

[Wispborne]

There may be more, but here are a few public things trained on David's art.

https://civitai.com/models/52480
https://civitai.com/models/30
https://civitai.com/models/8296
https://huggingface.co/Severian-Void/Starsector-Portraits

Three are from Severian Void, one posted by a Kiktamo.

HuggingFace is like Github for models, Civitai is like, idk, NexusMods. Both are extremely searchable by the public and indexed by Google.

I'm not sure if David is aware of this or not - I know it's not my fight, but *specifically* copying one artist's style and making it easily usable by anybody (rather than something that you probably wouldn't find unless you were part of this modding community, ie not indexed by Google) seems different from training a general-purpose model from hundreds or thousands of artists. Or maybe it isn't, and the only difference is that these websites have options to file takedown requests. Maybe you saw these months ago and have already had this discussion.

[Goumindong]

It understands the concept of a tree and guides the noise to something close to what it knows as a tree

So… this isn’t true. Wikipedia says that it lists “concepts”. But it does not. It has a set of correlation data that is tied to the bits and then tests the correlation until the images are sufficiently high. This is not the same as an understanding of the concept. It has a record of how other people have utilized that concept but does not know what a tree is.

While it’s true that the final model does not have the copies in it this does not matter. The correlation data is uniquely determined by the dataset and the dataset utilizes it as a fixed medium.

[Void Ganymede]

"This is not the same as an understanding of the concept" in the same way planes can't fly like birds. There might not be a legible constructive world model for a "tree" composed of "bark" and "rings" and "cellulose". That doesn't mean a world model doesn't exist - some basic experimentation will show that it *does* exist, it's just alien and illegible.

For nontechnical users the fight over anthropomorphism was lost before it even started. These are household spirits, not programs.

I know it's not my fight, but *specifically* copying one artist's style and making it easily usable by anybody (rather than something that you probably wouldn't find unless you were part of this modding community, ie not indexed by Google) seems different from training a general-purpose model from hundreds or thousands of artists. Or maybe it isn't, and the only difference is that these websites have options to file takedown requests. Maybe you saw these months ago and have already had this discussion.

How much of the fury is due to a lack of giving credit where it's due? Crediting inspirations, contributions, homages is pretty well-understood in art and software communities. It's weird that LoRAs didn't inherit that habit.

[MesoTroniK]

It's weird that LoRAs didn't inherit that habit.

Is it? If (most) TechBros cared about that sort of thing then (most) would not be using and making this sort of thing.