Examples using Photoshop’s new “Generative Fill” feature
twitter.com
twitter.com
https://github.com/Mikubill/sd-webui-controlnet/discussions/...
So far it seems that the OSS diffusion models + tooling that we can run locally keep being state of the art. It makes me so happy.
(Unlike say LLMs where GPT-4 is clearly dominating for now.)
Since we have capable local hardware, I'll propose this as an alternative once we get an estimate of our Firefly costs.
A lot of this comes from people comparing GPT-4 to e.g. LLaMA-7B, because that's the thing that fits in memory on their laptop. Whereas you can run LLaMA-65B, and it's dramatically better, but it uses about 128GB of RAM and the hardware needed to run it fast is expensive.
And GPT-4 has even more parameters than that, but that's not a matter of the tooling, it's that someone needs to release a public model with more parameters.
I can't do that with LLaMA-65B. (Although to be fair 128 GB RAM is not that much.) But I suspect it's still far less capable than GPT-4, is it not?
Then there are ones where more parameters matters.
Conversely, LLaMA doesn't say "I'm sorry Dave, I'm afraid I can't do that."
"....dropped...." in the context of new albums/music being released means "made available", however in the context of software features (or general english vernacular) it means "removed".
examples "When Weird Al's song 'White and Nerdy' dropped, I stopped everything so I could listen to it."
"In the latest news, Microsoft has dropped the ability to log on locally to your PC. All logins require internet connectivity and a MS Account."
This isn't a gripe about HN 'headline' rules; it's a complaint about a use/misuse of slang. Oh get off my lawn too please.
For example, Google widely advertises that its Pixel phones get big "feature drops" every few months.
Here are n [jawdropping|amazing|stunning] examples of what people has done with it.
These bullshit copy-paste threads from AI "influencers" and devrel hacks are a scourge - bandwagon "content" from people who produce nothing of value themselves other than tricking people into buying what I can only assume is $500 video courses repackaging 6 month old blog articles from someone else.
There are courses about "how to go viral on Twitter" that work, because the algo is so cookie-cutter and easily gamed. Back in 2022, it was the "collect list of useful links, add 5 of them to a tweet at a time, and make it a viral thread with that annoying finger pointing down emoji"
I think they are all created Nov 2022 or later, or were previously "here's a list" hustlers and pseudo entrepreneurs.
After seeing things go from 0 to Stable Diffusion and beyond, the fact that GenAI can do x,y,z; well, seems like a given eventuality.
- https://twitter.com/maddoxrules/status/1663715966755430401
https://www.reddit.com/r/ChatGPT/comments/13wfaqg/photoshop_...
https://share.getcloudapp.com/2NupkrOg
I remember when CS came out and it had the magnetic lasso tool, overnight people could save hours and hours of time in their work, this reminds me of that.
Before someone brings up future screen tech, there are additional reasons that's not a trivial workaround. Additionally there's a higher benefit to having high resolution cameras which have a wide field of view and the user may want to zoom in on the physical data vs a screen where the zooming happens in software.
Of course, nothing stops anyone from lying under oath. Police do it every day.
But I can see cameras with cell chips inside and auto-upload to an authentication service with the location verified by cellular triangulation.
A tangent but I think journalism is going to start to actually mean something. Historically I think it's been most problematic bullshit, and especially in the past few decades it's pure propaganda and greed and profit. But in an age where we can't trust literally anything we see, it's the reputation of specific humans we'll trust, and we'll have to trust that if they say they saw something, it was real. We'll start to see "real" journalism.
Previously, such manipulation was possible via manual methods requiring time and unique skill which limited it to highly motivated, organized adversaries and it was much more useful because most people weren't aware it could be done. IMHO, all the AI examples we're now seeing bemoaned in the media as signs of future manipulation are, in fact, rapidly diminishing the effective value of such deception.
I've fired up Midjourney a bunch of times and outside of fantasy landscapes, it has never managed to make anything useful in my actual work.
GPT-4 is increasingly going the same way too. I'm not sure if I'll renew the subscription.
Getting a little jaded of these hyped up announcements that are largely nothing more than cool tech demos.
Oh yeah, can't see how this is going to backfire at all...
Eventually, the cost benefit shifted greatly towards being less safe because of how much faster people can be. Anybody using AI to generate images is going to be way faster than someone not. Any company selling generative AI is going to be miles ahead of open source alternatives.
Anyone not embracing or figuring out how to bend the rules around assets is going to get smoked in a business.
It would be nice to at least try to get it right this time.
The open source community is doing more with GenAI and they're doing it better. If Gimp weren't in such bad shape, Photoshop would be over. This exact feature would already be ready to go and would be far superior to Adobe's version.
Any startup with minimal effort and capital will be able to duplicate this product. They'll eat Adobe's thick margins away.
Only been saying this for 15 years or so.
And yet, here we are.
Even Godot is going with that playbook like Blender and doing what GIMP never did. It’s a shame imo.
Imagine a world if GIMP and other open source creative tools had acted this way instead of the status quo of "Patches welcome" and "NOFIX", the attitude in the 21 year old "Adjustment Layer" feature request comes to mind...
but thinking you can fight adobe's photography setup (lightroom, ACR, photoshop, bridge etc) with minimal effort and capital to gain users for 6 bucks a month is delusional. but hey, if it happens, awesome.
Stable Diffusion is extremely cool and is a great project, but it does not compare favourably to Midjourney[1] or Adobe's offering. On the LLM side, StableLM is terrible compared to competitors. Some other "open source" models aren't open source at all, but instead were the "this doesn't compete so let's just dump it in the open" reactions.
Add that Adobe has a moat of legally licensed content that they used to train their model, which might turn into a massive differentiator if penalties fall hard on competitors.
[1] - A lot of people speculated that Midjourney used Stable Diffusion, but the products have veered so completely away from each other, the former just dramatically better for almost any prompt, that if they started with it they have made enormous improvements.
That said, OSS is in way better shape in the image space than in the LLM space. There are actually products and installers based on SD out there, something that won't happen in LLM space for awhile.
I'm not sure what this phrase means, can you explain?
Given the demos in the thread, I'm seeing a lot of 'production line' graphics being done a lot faster and easier with this feature. I'm not seeing a problem, but I'll eat my words if this will backfire like you said.
(I won't eat shoes or anything though, I'm not that confident!)
Junior loses job, company loses millions, potentially the client relationship, client loses perhaps even more. Everyone loses. All because it's not clear this feature sends your project files to the remote server.
No.
As the joke says: Patient: Doctor! Doctor! When I push here it hurts. Doctor: Well, then don't do it!
If you are working with potentially highly encumbered / lock and key assets then keep them on your computer. If you are worried that your employees might leak them, then block the connection with a firewall.
$20 says your inbox (man or woman) will be flooded.
Unfortunately, the posts have now been deleted, but you can still see other people's reactions:
https://www.reddit.com/r/photoshop/comments/13qtntz/generati...
With the Beatles it generates a psychedelic scene and an actual yellow submarine from what is really a very mundane street scene, indicating the model has 'knowledge' of who the Beatles are and what their aesthetic is.
With Master of Puppets, it utterly fails to determine the tone or even content of the scene.
Granted sure, if there was a serious lawsuit I'm sure going forward defense lawyers would do discovery to try to see if it was AI generated or not. But if it wasn't, then it'd be pretty trivial to show the original Photoshop (or whatever) files and where, if anywhere, AI was used and that it was just one part of a human created composition and that'd be that.
168h / 24h = 7d
I believe that words mean what people decide that they mean, and those meanings can change, but my perception is that this usage of "drop" has got a lorlt more momentum recently.
/s
https://helpx.adobe.com/photoshop/using/generative-fill.html
tl;dr: training data they were allowed to train on.
I PROMISE you once Adobe has market saturation on this and people have embedded it in their workflows and work without it feels painful they're going to be slapping a $40-$100 (It replaces literal days of work and literal team members, they can charge what they want) a month price tag on it on top of Creative Cloud.
Artists once again, never learning from the past and handing over their future to an abusive company that seeks rent on their ability to earn.
All you had to do is install any of the countless Stable Diffusion UIs and you'd have this tech free forever, some of them even integrate into Photoshop.
No sympathy for anyone who chooses this path when we finally had an opening to harm the Adobe monopoly.
Probably not what most pros are using, but perfectly acceptable tools.
The main advantage Adobe UI has is many people assume they’re stuck learning it at some point. But it’s definitely a waste of time and money if you’re happy as a hobbyist.
The hobbyist market is already served by Photopea, Canva, Snapseed and many others.
Think the artist equivalent of the guy building a AI controlled sentry gun for shooting squirrels with water.
Unless you're arguing that GIMP and Photoshop are treated as identical software by the market, image editing software can not be a commodity.
Alternatives to every baseline feature they currently charge a premium for exist, but because its industry standard everyone has to pay them.
Nothing Photoshop does today is special, in fact some of the ways it does things are worse than everyone else, e.g many filters being single core constrained and CPU bound.
> Nothing Photoshop does today is special
You seem to be falling into a fallacy common among technical people: that a product's only value is its technical implementation.
Photoshop is sold to businesses. It's value prop is "you will be able to hire anyone and they will be productive day one, which will save a lot of money in training."
People buying photos op don't care about filters being CPU bound. They care about turning work around quickly and getting the next paying gig. A CPU bound filter is statistical noise compared to having to figure out how some different tool works.
I work as a creative director, my anger about this is decades of suffering what I consider substandard tools.
> People buying photos op don't care about filters being CPU bound. They care about turning work around quickly
We definitely do care about our computing power being used to the fullest of it's potential. This is why some of us even do 2D static work in After Effects now because its faster and uses more of the machines power than PS.
I'm really curious why you care about using machines to their potential. Is it cost savings from not having to buy a more expensive machine, philosophical dislike or waste, or does it really come down to hours of non-billable time?
I think this is a little bit unfair and trivialises the problem. I know many artists who have to/had to use CC for their work because their clients or (mostly) employers consider it the industry standard.
I agree with the sentiment, fuck Adobe and decades of their parasitic practices, but the reality is not that black and white.
If not you can use something cheaper - or use one of the 3rd party plugins. If the workflows become popular, the 3rd party plugins will get better.
No kidding. I have only negative things to say about Adobe, but that is a terrible argument. $100 a month to replace multiple $100k/year employees? That's basically the definition of technological advancement freeing labor to be more productive elsewhere.
At the end of the day, not everyone will need to "make" generative art. There will be plenty of it on Adobe Stock that clients will be fine with, especially once they see the hourly billing rates for custom generative art that requires multiple prompts to get right.
You mean the ones trained off the artists art without permission, unlike Adobe's offering?
I know a lot of graphic designers that looked at that and hated it because of the fact that they took stuff without permission.
I don't know a single artist who was happy with that. The AI generation wasn't the issue. It was the training source.
I'm not an artist, and I'm also aware of the many failings of Adobe, but you really missed what the issue was.
So, if an artist's work makes 100$ per work an AI system which can generate the same quality work but 1,000,000 pieces a day this won't end up the AI system generating 100 million dollars a day.
It will simply rise the bar for artists instead. Which is a good thing for the society as a whole because the baseline will be very good and the work of the talented artists will be highly valued.
Adobe can eat the market of mediocre artists if they are the only one with this technology so they can control the output rate and create artificial scarcity. They are not the only one with this technology.
Obviously AI generated art has market value to companies with creative staff because AI is designed to devalue and commoditize the work of actual artists, such that the work of all artists can be replicated by someone on Mechanical Turk writing prompts for pennies a day (or some equivelent low-wage plebian drone.)
AI has already been used to steal commissions from working artists, and companies are already firing their entire creative teams to lean entirely into AI generation. The better the technology gets, the worse the situation gets for working artists because that's the plan.
I don't see how anyone can have witnessed the glee with which AI techbros danced on the graves of the art industry, ridiculing and harassing artists, calling them gatekeepers and fascists, and telling them their careers were doomed, and think this will benefit artists in any way.
I really want that to change, because I find this technology fascinating, but as of now the well has been poisoned, the bridges burned, war declared, and most artists with actual talent would rather starve in the street than touch anything AI.
Then you have the artists who matter, whose work is valued because it has their name on it, regardless of what process was used to create it. AI will be just another tool they may choose to use.
"struggled to get paid for expensive manual labor?" You seem to have a contempt for most working artists, because most struggle to get paid regardless of talent, not because they're incompetent but because art of any kind has always been a difficult market. And yes, it's hard work, which many people fail to appreciate. You're really illustrating my point more than I think you intend to.
Perhaps we have a very different understanding of what "art that matters" is, certainly I'm not thinking of anyone producing art at volume for a corporate IP. In any event, nothing ever stopped such a corporation from reproducing a certain style with cheaper labor, whether it's human or not. If you were, say, an art director working for a corporation, your job was never to perform much of that manual labor. AI isn't replacing you.
Moreover, the emergence of mass manufacturing always comes with a counter-movement attempting to maintain the "human touch" and charging a premium for it. For every Budweiser, there's a thousand craft beers. Neither replaces the other.
As an artist (creator) myself, I do not feel threatened. I do not consider the tedium to add anything of value to my work… and hell, it’s suck to integrate /generate fill (as an example)
It seems like generative AI will wipe out many design shops and non-IC/exec brokers/managers of creatives (recruiters, talent agents, admins, etc.) at least to the degree the web wiped out newspapers and their non-creative staff.
But it'll also enable creative designers/artists/writers to become their own studios. As Kubrick said, "one [person] writes a novel, one [person] writes a symphony." Generative AI will enable one person (or a very small team) to create a blockbuster movie or AAA videogame.
Perhaps. But what would that change and do you have any choice? 'Techbros' can make do with genAI or even stock images. I'm not so sure about said artists.
AI can generate an image "in the style of Banksy" but so can a thousand other artists who aren't Banksy, and their work product would have no more value than the AI's.
The way this makes you feel is about more than texture and shading:
https://www.sothebys.com/en/buy/auction/2021/contemporary-ar...
You might be able to get Stable Diffusion to generate something like that with a sufficiently detailed prompt, but then its uniqueness would come from the creative effort required in devising the prompt. It's not like you can just type "street art in the style of Banksy" and expect the output to make you feel something like that.
I Agree, those who produce the training data or the thing that the AIs work will be based on, should be properly compensated.
Being serious so, human art is what it is because it requires effort. AI art doesn't require nearly as much, is generic and abundant. In short, boring.
Remember:
Duct tape + banana + wall + 3 minutes = incredibly valuable and controversial art.
Me taking a year to paint a detailed depiction of a bowl of fruit = worthless, terrible art.
Results are what people care about, not effort.
There is a sci-fi prediction that humans would eventually reduce to a thing with an eye and a finger that only presses a button, and machines would do the rest; I take it as an optimistic view that it means humans remain relevant, and that human values only come from humans.
But Photoshop's generative fill is trained on Adobe's own stock photos, which removes an entire class of potential future litigation issues.
If we move to this lie that Adobes models are "ethical" it gives them a moat to make free models illegal so Adobe becomes the only option.
Artists should be realizing the goal is to stop Adobe becoming the only option, if they do become the only legal way to do this then expect AutoCAD level pricing.
The alternative is to not pay for things, or?
First law: energy can neither be created nor destroyed
Second law: entropy always increases
> There's a reason the first law of thermodynamics is first
The first law of thermodynamics is not the first. That would be the zeroth law:
Your average user doesn't understand how to set up and leverage open source software because it's too complicated to get going.
They want to open Photoshop and use the feature in the tool the already use, without setting up anything.
Quite frankly, Adobe nailed the UX here.
I expect most of us (hackers) have tried these tools, but not your average user. When I see comments like OP, it's coming from the hacker perspective imo.
The user is not making an error, the error is assuming everyone should be a hacker.
It gets worse, there was a thread a couple weeks ago about Envato (ex ThemeForest) asking all their authors for "free/unlimited" rights for anything they have to be used to train AI models for NO compensation. It was eye opening how many "artists" and "producers" accepts the status quo as is, and think it's still a good idea to collect whatever this company throw on their way.
It's really hard to feel sorry for these people...
The mistake companies make is in not fully embracing this.
Alright, once you get it running the UI for generative fill can be used by a graphic artist, and now you're several months ahead of your competitors, but still the UI is a little rough. But it's open source, so you have your techs make it a little better.
Now your company needs to make it easy to submit a pull request, or you're suddenly maintaining your own fork, which you're going to want out of as soon as you have an alternative, and soon your company is paying $$$$ to Adobe again.
Isn't the point of your UX to be the best available?
Implying malicious intent to designing good software is a bit weird, whatever Adobe might be doing in terms of delivery and pricing.
If MSFT and Apple each shipped a Dropbox competitor that was highly integrated, available, and pushed flagship software suits to adopt it as well within a year, would Dropbox had become as large as it did? I would guess not. It would likely be a downgrade from what Dropbox offered (the separate ecosystems staying bifurcated comes to mind), but a zero install solution would be the only thing that could make it easier.
Most designers work entirely within software like photoshop or figma that are auto installed or cloud based.
too many college canteen marxists in this place is my thinking.
That is such a strange thing to read. I'm not contradicting that it's true in this case, but Basically every artist complains about how much of a clusterfuck the UI/UX is while being entirely happy that the UI is horrible enough that intimate knowledge of it keeps them with a job
Its much more capable than the alternatives I've tried. But the learning curve is steep.
The productivity gains are becoming permission-to-play. The money saved will go into hiring different kinds of people which will continue to pressure the software budget. It’s true CC is seen as necessary, but it’s still resented and there are always employees whose access to a subscription is marginal, providing some elasticity.
I think 100 a month is a low price for this.... if it saves literally man-days of work. The equivalent value addition is something in the ballpark of 3k per project (assume 100 an hour over 3 8 hour days).
Add in the fact that your velocity has increased significantly, they could probably get away with at 5k a month at the "Pro Tier"
Do you have any specific suggestions you favor?
https://twitter.com/search?q=photoshop%20generative%20fill%2...