These posts are all written with this insane premise that the moment in time the post was published, will continue to be the status quo for more than a couple months.
> If you’ve ever used a generative system, I can pretty much guarantee that you spent an embarrassing amount of time making tiny adjustments to your prompt and retrying. Producing a compelling image with generative AI is pretty easy; maybe one in ten images it generates will make you say, “Wow, cool!” But producing a specific image with generative AI is sometimes almost impossible.
Who could possibly think this will be the case six months from now? I mean, maybe some of the the content and warnings here is fleetingly accurate, but it's truly not a hill to die on. You could fire your illustrator and be inconvenienced for a couple months until the next stablediffusion update. It's a disservice to illustrators to make them feel safe.
I think firing your illustrator to replace them with Midjourney is like firing your developers and replacing them with Copilot. You still need someone to do the job, but now you got rid of the person who knew whether the result is actually good. We might as well get rid of photographers and illustrate our news with cell phone pictures while we're at it.
If a company needs to generate illustrations at a big enough pace to require an employee they'd only be replacing their illustrator with a worse one. We all know what happens to GUIs when programmers develop them, so why would this case be any different?
> We might as well get rid of photographers and illustrate our news with cell phone pictures while we're at it.
Wait... Hasn't this already happened? Professional/specialized press photography seems way down compared to pre-smartphone era? Now a journalist/reporter is expected to do a passable photo job on their own. Or a random member of the public
I'd bet on this being true six months from now. DALL-E 2 came out in 2021-01 (19mo ago) and Midjourney in 2022-07 (13mo ago). While the quality of images has improved a lot, if you have something specific and moderately complex in mind it's still very hard to get there just from a textual prompt.
> While the quality of images has improved a lot, if you have something specific and moderately complex in mind it's still very hard to get there just from a textual prompt.
I think this gets to an important point. If whoever is paying the bills has something very specific in mind, they won't be happy with AI (or frankly many artists) at this point. But as a creative interpretation of something more general, it's actually really good, and I think with many low-importance works like the author describes as "furniture" - we really don't need to be that exact.
One year ago we had textual inversion. Now we have LoRA and control net. I know it might not be a scientific breakthrough, but in practice it's day and night.
I'd be happy to bet that the tech will continue to improve. But "producing a specific image with generative AI is sometimes almost impossible" seems very likely to still be true in February 2024.
Of course an AI can not produce an image that matches what's on your mind 100%. Because by definition it will require you to provide all the information, a.k.a. you need to make the image by yourself first.
But I really don't think illustrators are as safe as the article implies. Yes, the jobs won't disappear overnight, but are there that much demands for illustration to support a future where every illustrator becomes 10x more productive than before?
(I've done illustration commercially before, while it's not my main source of income and I'm junior level at best.)
I was kinda tempted, but I don't think there is such a good metric to measure it. (Since the job market is heavily influenced by interest rate etc... and if I were confident at predicting macro economics I would just bet on stock)
What if we picked 10 prompts that it seems like an AI should be able to depict well, but it can't yet? And then iff the best AI tool in February can do the majority of them you win?
I'll bet against it tho. I don't actually believe pure text-prompt-to-image will improve much (not in a few months at least). I just believe there will be more non-text tools to guide AI, like LoRA and control net, and they will be more accessible.
Control net kinda did what you said but in a different timeframe: it was quite difficult to tell AI to generate a person "sitting with their legs cross". Today, it's relatively easy to do this with control net, but still hard with text prompt only.
Edit: and the sibling comment made me question myself why I would ever take a random bet on the internet.
I was just stating the improvement on SD we've seen since DELL-E 2 and Midjourney came out wasn't just about "quality of image", but also about "have something specific and moderately complex in mind". Thus I mentioned textual inversion vs LoRA/ControlNet.
I wonder why people take random internet bets, I've seen this on Twitter as well from some prominent people in the tech world, who just like betting on outcomes I guess. I saw it most recently with the LK-99 "is it a real room temp superconductor or not?" bets.
I think often someone tries to frame things into bets to pierce the layers of instinctual contrarianism and "have strong opinions weakly held" that often pervades internet discussion.
More charitably, reframing into a bet gives a relatively neutral opportunity of re-stating the discussion/argument to more clearly identify areas of agreement/disagreement (since a clear definition of the disagreement is crucial for forming a bet).
I don't think it will be true six months from now. OpenAI has been red teaming a new version for months that goes way beyond anything we've seen today. You can see some leaks from this video: https://youtu.be/koR1_JBe2j0
I don't know if that's 'way beyond'. Those samples look similar to Imagen and Parti in terms of quality and following complicated text prompts. (Look at the PartiPrompts for the absurd prompts Parti can execute accurately.)
I'm glad to see OA does have a successor for DALL-E 2, though: the service seems to have somehow gotten worse since release.
I shouldn't have said "goes way beyond anything we've seen today", but probably it goes way beyond anything we will be able to try in this 6 months period. Parti 20B is really good at language understanding, more meh Imagen.
>the service seems to have somehow gotten worse since release.
Yeah, I think it's pretty clear that they probably lowered the number of steps or used a similar strategy to reduce the computation.
But in general the better something is the harder it is to improve.
From an outsiders' perspective it looks like we've had a series of achievements where AI worked impressively well considering it isn't a human. But afterwards we got an incremental grind that never got close to an AI that's just good, period.
I'm very impressed people got self-driving car demos working. I couldn't do that. But year after year self-driving cars remained a demo, albeit an incrementally improving one.
It's limited to a small number of cities with optimal conditions and every ride has a human supervising. They're still spending R&D, rather than making a profit. That's a demo.
> It's a disservice to illustrators to make them feel safe.
It is also a disservice to encourage managers to make their illustrators redundant when we don't yet know that, in six months, AI image generators will equal a human illustrator in being able to create adequate pictures. This kind of article has an important role, namely that of countering hype (often based on PR from machine learning companies). That hype is why more ignorant people think that AI can already replace crowds of employees, when the reality is more nuanced.
Why are AI people always trying to convince me to do something now based on an argument that it will work later? Regardless of whether you're right a working model six months from now is not a working model now. Wouldn't your hypothetical person be much better off firing their illustrator in six months?
Of course. But what should the illustrator do today? Considering that there might be a higher risk of getting fired in N months/years from now. Probably refine their business skills. Maybe also get familiar with the capabilities/weaknesses of the latest tools, and how they can use them to be more competitive than before?
You have total doomers aware of the AI potential. Horrified at the long term consequences of these LLMs. Then on the other side you have users saying "Its just another crypto bubble", and when pressed, they admit that they never used it.
There is just such a vocal population here that says 'Well its not always 100% perfect, so its useless", and they are burying their head in the sand that companies are already using the OpenAI API to reduce the cost of business.
I genuinely don't understand these people. They don't use the technology and they deny how useful it is. There is news and real world examples of its usefulness. I can only imagine these people manage (money) terribly.
History tells us that both the doomers and naysayers are probably wrong.
It seems to be pretty solidly demonstrated now to have some limited efficacy across a broad range of areas today, and is very effective in some niches (like the articles mention of producing SEO fodder cheaply).
Growth from that state though? The only thing you can bank on is that nobody really knows.
I don't understand people who think an imperfect product provides linearly less value when it obviously provides exponentially less in most use cases.
If ChatGPT can replace a team of software engineers why didn't you replace that team with four times as many college interns years ago? Because you can't combine people capable of doing easy coding and get someone who can do moderately hard things.
This is more or less my opinion as well. This is a list of current weaknesses of generative AI. No, right now I cannot replace my graphic designers. I bet I can in a few years, though.
> If you’ve ever used a generative system, I can pretty much guarantee that you spent an embarrassing amount of time making tiny adjustments to your prompt and retrying. Producing a compelling image with generative AI is pretty easy; maybe one in ten images it generates will make you say, “Wow, cool!” But producing a specific image with generative AI is sometimes almost impossible.
Who could possibly think this will be the case six months from now? I mean, maybe some of the the content and warnings here is fleetingly accurate, but it's truly not a hill to die on. You could fire your illustrator and be inconvenienced for a couple months until the next stablediffusion update. It's a disservice to illustrators to make them feel safe.