Also, server/hardware costs are still a limiting factor for running and finetuning the larger 33/65B Llama models. Especially, if they can only be used for personal toy projects.
Midjourney is way more versatile than SD. If you start getting some fine tuned models on civitai, trained to do well some specific tasks, you can get comparable quality but I haven't seen a single model which is able to replace Midjourney.
Llama is no different, it has ok performance on generic queries but still far away from GPT3.5: if you start fine-tuning you can get good perf on specific tasks.
Sure, its very easy to get good results fast, but the tuning that avoids "uglier" images is the same that removes a lot of versatility compared to SD
Also controlnet is a killer feature
They’re very similar offerings if you’re willing to put in the work on SD.
Sure, Midjourney is a centralized commercial service with a clear statement that you (as a paid user) own the images you create. While that doesn’t resolve all potential copyright issues (as there are still at least theoretical issues with the underlying dataset), if you doing something commercial with it like, say, a webcomic from which you derive income, its a lot simpler than dealing with the SD ecosystem where the plethora of models also have different stated usage restrictions, different suppliers (many of which are hobbyists) to keep track of, and more potential avenues of indirect copyright risk, as well. For some webcomics, even the base CreativeML Open RAIL-M license itself might be problematic.
This isn’t a technical or quality advantage, but its definitely an advantage that would very often tip the balance between two tools if both are minimally adequate to your task.
For some technical reason?
https://www.gesetze-im-internet.de/urhg/__44b.html
https://www.gesetze-im-internet.de/englisch_urhg/englisch_ur...
To push a bit further, there’s something that just feels particularly off about assuming everyone’s content is up for grabs unless the producers do the work to opt out. I think there’s an especially palpable bit of irony looking at it from the EU’s perspective—where cookies must be opt-in, but grabbing all your copyrighted material so companies can do whatever they like with it places the burden on the owner to opt-out. It just feels backward. Perhaps one should have to expressly opt-in to allowing their work to be accessible as training data. At least then there will be a clear signal that the producer of the work can’t later complain, as they willingly granted permission.
While you don't need that to generate an image, it's something SD can actual do extremely well with ControlNet, Textual Inversion, LoRA, img2img and so on.
That's an area where things are going to get interesting in the future, as you can take any image, feed it into SD and produce hundreds of AI images from it. Very easily, without much effort and within minutes. The delineating line between original work and derivative becomes extremely blurry here, as what you are copying is not "the image", but just concepts within the image, that can be a pose, camera angle, scene layout, art style or really anything. You can "copy" it with as much variation as you want, you can remix it with other images, text prompts and so on. Where does "looking at reference" stop and "doing a copyright violation" start?
The spooky part with AI art that it stops images from being singular entities, with AI you can explode every piece into millions of possible variations. AI is so fast at generating art that a future where we could generate movies in real time might not be far away. It's already fast enough to produce images and text stories faster than you can consume them. There might be a fundamental shift in art consumption ahead of us.
"A reservation of use in the case of works which are available online is effective only if it is made in a machine-readable format."
It's also far from clear to me whether a court would find training an LLM to constitute text and data mining 'for the purpose of gathering information, in particular regarding patterns, trends and correlations'.
It's pretty normal for a law to not be specific on the technicalities so they don't have to update the law whenever the software changes. The de facto standard to prevent bots from scraping your sites has been robots.txt for almost 30 years.
If artists didn't mind Google scraping their images, putting them on their site, adding ads and making billions, I really don't see them having much of a justification to call out StableDiffusion for "stealing" their stuff. In general artists would be in a lot of trouble if taking stuff from the Internet would be outlawed, as that's where they get all their reference images from too.
Either way, I am sure we'll see quite a few lawsuits going forward, laws are always open to interpretation, especially when new technology archives. But long term I really see copyright in general being in a lot of trouble, since derivatives and remixes are becoming completely trivial with AI. Where does the original work stop and the copyright violation starts is being rather difficult to decide when you can just wander around latent space and create literally thousands of similar images in minutes, with as much or as little variation as you want.
While a court would likely conclude that a watermark on an image is not 'machine-readable' (I say likely—OCR technology would however make it possible that a court could find that a watermark is machine readable), I would say that because the law does not require a specific method, I think it might be found that a copyright notice in the footer, or in an image caption, is indeed 'machine-readable'.
On balance, I agree that there's a lot of things we are woefully underprepared for coming up in the very near future on using tools in this way to generate art. The answer is not simply to try and lock up all the art away from the robots—but I don't know what the answer actually is.
I think you're being a bit generous there. Either I'm using it seriously wrong or SD can only generate vague blobs while Midjourney can make some proper stuff. It's a larger difference than GPT 3.5 vs GPT 4.
You are definitely using it wrong, if the alternative is “SD can only generate vague blobs”. Even the base SD models are much better than that (though, the strength of the SD ecosystem is the availability of custom checkpoints, hypernetworks, LORAs, embdeddings, ControlNet, etc., not just the base models.)
Not sure why there's even an option to go below 512.
Models on CivitAI are okay. Cool if you're looking for a certain style and/or want to create something that looks like the training images but style isn't everything.
Midjourney generates much better than "slightly better images" and the very fact you say this just tells me you've not even used the thing in any real capacity.
I am the author of submissions such as: https://news.ycombinator.com/item?id=35181433, and I am one of the people responsible for the enthusiasm behind the performance of MJ v5.
But no, MJ is not much better if you know how to use SD, although if what you did with SD was just put a prompt in a huggingface space, I can understand why you say that.
>I never said you can't do impressive things with SD, but feel free to share these comics.
I am arguing that they are better than any comics made with MJ, not that they are simply impressive, that's really the entire point. I know some on Pixiv, you can look them up if you want; I am not linking them for obvious reasons (to say they are NSFW is putting it mildly).
https://www.pixiv.net/en/artworks/107271972
Though, if they're training off official character art that's less cool than reinterpreting it themselves. Means you don't have a "house style".
I'm the person behind these - [url-redacted]
I think it's safe to say i know something about SD's capabilities.
>I am arguing that they are better than any comics made with MJ, not that they are simply impressive, that's really the entire point.
Sure that's why i'm asking you to link these comics that are supposedly better than anything Midjourney has ever produced. With a claim like that, i'm sure you understand wanting to see results.
>You can go look them up on Pixiv if you want, they host some; I am not linking them for obvious reasons (to say they are NSFW is putting it mildly).
So you can't link anything that isn't NSFW on pixiv? Lol, that just solidifies my point. Frankly if the best you can come up with is pseudo porn(or maybe not pseudo lol) on pixiv (i don't imagine any readers of that will care about the things i'm looking for) then that's not a very good look.