Adobe Firefly vs. MidJourney AI image generation
twitter.com
twitter.com
"Same prompt as above but for Pikachu. Again, somehow Firefly does not fully get these famous characters. Maybe a training data copyright issue?"
Comment in hacker news thread some days ago https://news.ycombinator.com/item?id=35247720 :
"Adobe employee, not in Creative Cloud. [...] The team is going to great lengths to make this ethical and fair (try and generate a photo of a copyrighted character like Hello Kitty or Darth Vader)."
I'm an artist myself, absolutely love Midjourney, and have absolutely no problem with it being trained on any and all art -- just like human artists have been doing for millenia.
https://en.wikipedia.org/wiki/Copyright_protection_for_ficti...
But in both cases, artists are particularly well protected in making artistic commentary.
Artists' tools should permit fair use, and disagreement can be directed at the artist.
Until that is settled, its likely that Adobe is taking the "we're not going to deal with any licensed properties in the model to avoid any lawsuits."
https://vox.com/recode/23580554/generative-ai-chatgpt-openai...
It depends what you call "serious" but on their government relations page https://www.adobe.com/about-adobe/government-relations.html they say "Adobe discloses information on our federal and state lobbying activities. Our most recent federal lobbying disclosures are listed on the U.S. Senate lobbying disclosure database and U.S. House lobbying disclosure database."
If you look at the forms on those databases you can see that Adobe has a "Policy, Government, and Ethical Innovation team in Washington, D.C." which I guess is what they call their lobbying crew.
> It's Disney you've got to watch out for
Probably most lobbying efforts would be considered 'not serious' if the standard is Disney!
If you used a really cleverly and exhaustively and iteratively prompted midjourney image to recreate a copyrighted painting, that WOULD be duplication, just like a physical artist can paint a copy of a copyrighted painting by hand.
Duplication is the end result, not the intermediary. That's why transformative collage is fine.
if you turn the temperature value down to zero it will produce the same thing every time given the same input (in the same way zlib or jpeg will)
it's a lossy compression function that is implemented using a neural net
copyright isn't removed if I take someone's existing image and re-save it as a jpg
if the courts agree and this is ruled to be infringement then we'll left with products with level of quality we see here
meaning the technology will be an economic dead end
here's to hoping
I don't think, given that we don't even have a particularly full understanding of how human conceptual logic functions, that we can claim that even AIs are using it as well. It's only "abstracting" in the sense that it has labeled one million objects in its training set with the word "tree," and fuses many of those images together to form a general picture dependent on specific parameters made to limit its set (oak tree, winter tree, etc.)
But that is different from me or you using the word tree, which is just a signifier among signifiers, it stands for nothing but a negation of the very thing it points to in a certain set of symbolic relations. Humans communicate in the order of symbolic structures, our minds function much more like LLMs, creating multitudinous pattern relations. What you call "abstract concepts" are of a secondary order imposed to create rigorous exactitude overtop the riddled mess that we call the human psyche.
What the model does is turn the picture into numbers in a clever way (embedding).
Then, the model compares the relationships between all the embedded tokens (self-attention).
Then the model does the fancy linear algebra stuff we've come to know and love (feed forward neural network).
What you're basically getting at with your temperature comment is "rather than having it semi-randomly pick one of the top n options that the feed-forward network recommends for which pixel goes where, we can always have the model simply choose the top answer each time."
That is not "we can make it reproduce training images"
The point of embedding is to see multiple examples of similar images and generalize from them. The process in and of itself breaks the ability to re-create a photo, unless the training set is very skewed.
If that's the argument of legal cases, they are likely to fail or win on ignorance.
didn't say that, all I said is it's a deterministic lossy compression function of the inputs (unless randomness is deliberately introduced)
and as such it should be treated the same way as zlib or jpeg for copyright purposes
For example, switching training order of images has a known impact on outputs, but I don't see the equivalent in traditional compression.
I agree with you that it is computer code that runs on circuits (and is therefore deterministic in a very loose sense), but it is fundamentally incorrect to compare it to traditional image compression.
This would only be true if it didn't mix or 'forget' any of the training data for each prompt, and only then for the exact same inputs of the training data. That is a lot of qualifiers that aren't practically applicable.
Slight side note: a deterministic lossy compression function should not always be treated the same as jpeg/zlib, as hash functions are technically also in that category.
I wouldn't want to be the defense's lawyer when that image gets shown to a jury.
https://www.theverge.com/2023/2/6/23587393/ai-art-copyright-...
Making absurd arguments is fun!
(The reason this is absurd is the number of parameters in the model is incredibly small compared to the size of the training set. You wouldn't even have enough bits per training image to keep a unique identifier for each training image, let alone a compressed representation. And to the degree there's redundancy in the training set, this is undesirable, actually counter-productive to the functioning of the model.)
(In a way, the human brain is actually more capable of copyright infringement since the number of parameters in the human brain is about 4-5 times that of a typical image generative AI neural net, and we're trained on a smaller and less generalizable data set. Humans can and do make copies of things from memory, which generative AI massively struggles with and can only do for corner cases.)
https://en.wikipedia.org/wiki/Copyright_protection_for_ficti...
Arguably, however, (a) not all fictional characters are copyrightable in the first place, and (b) fair use still exists.
Adobe Illustrator should not prevent an artist from drawing Mickey Mouse using the software. Nor should MidJourney.
If Disney has a problem, they can go after the human using the tool.
If you look up "Pikachu" in Adobe Stock [1] there actually are a lot of results. They're mostly figurines, pokemon cards, quokka, and Machu Pichu. Very few direct illustrations. And of course there might be filtering used on the dataset before it was used to train the model, a lot of the images that have Pokemon imagery are marked "Editorial use only".
1: https://stock.adobe.com/search?filters%5Bcontent_type%3Aphot...
Presumably, the dataset for training the embedding layer (which is trained separately and then fixed) is not stripped of copyrighted content. So it will have learned that "Pikachu" is related to words such as "yellow" and "rat".
Therefore, even if the image dataset didn't have a single picture of Pikachu, the image generator will still likely produce a yellow rat. Just one that doesn't actually resemble Pikachu.
I suspect there are a lot of the characters in the training data, but (postulating) there might be a post-generation adversarial selection step where anything that is recognizable as a someone else's IP gets filtered out of the results, so anything sufficiently Pikachu-like won't be shown, leaving behind the vague yellow monsters.
Such a system would be more robust against incidental infringement of IP not mentioned in the prompt. e.g. A prompt for "Theme park" won't create Mickey Mouse in the background
The entrenched will argue for regulation and get it. The entrenched that built everything on data they didn't actually own, will come out in favor of preventing that from happening again.
Here's the human figure (number 7): https://www.bing.com/images/create/editorial-style-photo2c-m...
Pretty solid Pikachu: https://www.bing.com/images/create/a-photograph-capturing-pi...
It had a hard time with the woman in red, but did better than Firefly: https://www.bing.com/images/create/modern-street-style-photo...
Totally flubbed the app icon, surprisingly: https://www.bing.com/images/create/app-icon-design3a-ios2c-s...
A little low on detail on the scifi city: https://www.bing.com/images/create/abstract-fractal-circular...
Here's a 'Soviet style propaganda poster of a female cosmonaut, illustrated, illustration': https://i.imgur.com/BEGHUsx.jpg
The same thing in Stable Diffusion: https://i.imgur.com/4Yfk7fO.jpeg
You'll note I tried to add illustration on to the end after illustrated to try and force it. That's because when it does this it tends to break the overall artstyle.
Another prompt I did was a dwarf carrying a rusty axe, fantasy art, etc. It worked fine when it generated white dwarves, but when it had races injected their faces were like photographs on top of rest of the art. One was a black man with dreadlocks and the other a Muslim woman with a headscarf. You can add "white" to the prompt to try and fix this but chances are it'll apply "white" to other things as well.
Also, their censor is hilarious. "Photo of a man in prison" gets rejected. "Photo of a woman in prison" works. You can't use the word attack, battle, World War 2, etc. You can't use the word "lingerie" or "babydoll" but you can absolutely do a "sheer lace cami set." That means this prompt totally worked when I tried it - "A 14 year old girl in bed with a sheer lace cami set," and the results were what you'd expect - sexy pose and all.
I honestly don't know how larger companies are going to deal with trying to censor this stuff. It's a really hard problem. I'm sure there's a combination of words out there that would let someone generate child porn. Potential legal issues aside I'm not sure Adobe has the stomach to deal with the Twitter/media fallout.
* Pornographic material and explicit nudity * Hateful or highly offensive content that attacks or dehumanizes a group based on race, ethnicity, national origin, religion, serious disease or disability, gender, age, or sexual orientation * Graphic violence or gore * The promotion, glorification, or threats of violence * Illegal activities or goods * Self-harm or the promotion of self-harm * Depictions of nude minors or minors in a sexual manner * Promotion of terrorism or violent extremism * Dissemination of misleading, fraudulent, or deceptive content that could lead to real-world harm * Personal or private information of others (like full name, home address, phone number, email address, government issued IDs, or anything else that would let someone locate or contact you in the real world). * Please note that we may report any material exploiting minors to the National Center of Missing & Exploited Children (NCMEC). If at any time you believe someone has violated these Guidelines or the Adobe Firefly Beta Terms of Use, please report it by contacting us at abuse@adobe.com.
To be fair my prompt didn't explicitly say anything inherently sexual, and I don't think the images would pass the test for child porn.
I do find that language funny - material exploiting minors - who was exploited? I'm waiting for journalists to figure out that it's a thing easily done in Stable Diffusion and probably less easily done in other image generators to see the outrage. Personally I think that if pedophiles could generate images that might mean there'd be less of a market for real images and therefore less (real) child exploitation, but I doubt Adobe will want to make that argument ;)
I hate to break it to you but as long as a piece of art is produced without sexual exploitation of children, even if it depicts children in sexual circumstances, it is a legal expression of the first amendment (in the US). Think of "loli" hentai, the myriad erotica stories featuring young children, even books you probably read as an adolescent depicting kids your age in sexual situations. A computer-generated pornographic lifelike image of a nude 13 yr old is completely legal so long as it doesn't depict a real 13 yr old.
The question is whether it is possible to do that, and if any pornographic images of children generated by these machines are not in some way eroticizing images of children in the (adobe) photo stock, thereby exploiting them. Either way, the cat's out of the bag now, as soon as the originals are lost and people are just using copies of copies, it will be impossible to track and therefore there would be no legal avenue to stop the proliferation of this type of pornography. It may even become difficult or impossible to differentiate between real child porn and computer generated child porn, and that would make it difficult to prosecute those who produce child pornography unless some new, very specific laws are passed in the US which avoid violating the first amendment.
At the same time, it also seems to be that a filter at the end, that analyses the generated image to check if it's really something you want to show, could be a good idea. But filtering and altering the prompt seems like a unreliable way to do it.
However, right now tuned to be super diverse and either it's injecting something like "a photo of a black woman from Atlanta" or there's something up with their model where if it doesn't match what would normally be done with the rest of the prompt it throws it off and makes it go from illustration/painting/whatever to a photo.
After I wrote my comment I tried running the same prompt on Bing and I think they actually do seem to check the image after generation, especially after you ask it for something it doesn't want to do. After I ran the prompt I did in my parent comment and it refused, I then asked it to just give me a photo of a "14 year old girl in a dress." It made the image but didn't show it to me.
Even that won't be perfect though, along with filtering. I have a Stable Diffusion model of my wife and I used nude and boudoir-y images in the training and it got a bit overtrained. Her in Stormtrooper armor? It's now sexy Stormtrooper armor. It once tried to generate an image of her as a baby/small child even though that wasn't at all in the prompt. I stopped it mid-generation but I'm guessing it would have been in some way "sexy." With diffusion models that stuff can just...happen.
However, I believe it is unfair to hold it to such high standards or compare it to Midjourney. It is new beta software and other AI image generating products have had this same problem and managed to resolve it in less than a year. I'm certain Adobe will keep working on the Firefly image pipeline until it produces decent results.
Firefly is absolutely underwhelming. Its been 10 months since Dalle 2, and it is just a slightly improved version of Dalle2 (With just as bad censorship filters). Adobe clearly only started their generative AI project only after Dalle 2 was released. Otherwise you would expect their internal team to have far more polished results.
Like, who exactly is a user of firefly at this point? You'd work far faster just copy pasting images from Midjourney into photoshop, than trying to wrangle firefly to produce good results. If you want strong control, SD is the answer. If you hate AI, you won't use it at all. Adobe will barely get feedback data.
This is not guaranteed, since Adobe has a much smaller library to call on for training data than the rest of the market. This can't really change unless they follow their competitors and ignore licences etc.
My hope with it is that despite mediocre image generation, Adobe will lean into their strengths and build really great tooling around it. I imagine a really great user interface around generation, inpainting, controlnet, image2image, etc - a UI that a well-heeled company has put a lot of money into designing and testing. I'd be ok with last year's generation quality in exchange for that.
So far, though, Firefly is just a demo of a few basic, outdated generation tools. I don't see any value in it yet, but hopefully it's early days for Adobe.
For example, during testing, I could not use the word "evil" or "gang", a number of celebrities and politicians, and a random assortment of IP (Star Trek related IP seems to work okay, but "lightsaber" does not and it's not clear if that's a Disney issue or a weapons issue, since "gun" and "pistol" don't work but "blaster," "cannon", and "knife" work fine.
The one thing I really liked from firefly was the text creation tool, it's really cool (although I have no clue what the applications are for people not working in gfx design)
Also, Dalle2 is from a year ago. Midjourney has undergone 3 massive version upgrades since then. Midjourney v5 completely crushes the DALLE and firefly on every perceivable axis.
MJ is on discord, because they need community moderation by forcing public generation. Its the most effective way to suppress deepfake generation given their model very dangerous if use unrestricted.
BTW, faces are horrible and unusable. Full stop.
Which is sad! They shouldn't have released this until it integrated with your existing Adobe tools and your own art.
Also... it "watermarks" your images with a huge ugly gradient in the corner, and it keeps track of all them with some metadata that's stored on a "verification" website, along with your name and tool's info. Just so you know.
Referring to: https://news.ycombinator.com/item?id=35418999
https://www.techdirt.com/2023/03/31/midjourney-ceo-says-poli...
I wouldn't want them as vectors fonts necessarily. One of my complaints was that duplicate letters would just re-use the texture. I get that it's probably a caching mechanism to finish generating the word faster. But, since the AI can generate infinitely, it seems a bit self-limiting to re-use a letter. I guess my complaint is that the re-use really points out that it's _not_ a custom illustration.
Without knowing that, it’s hard to compare.
It would be good to watch a side by side video of someone doing their creative process, using each of the tools, with the same goal.
Then you can see the reality behind generating those images.
How long it takes, how many tries, tactics used, etc.
After all, directives like "Warm glow, elegance" or "skeuomorphic style" are hacks that people have learned to use to specifically improve MidJourney's results.