Almost everyone who perpetuates this panic seems to be talking about prompting. Typically, that's how you spot people who either:
- didn't spend much time around those ML-powered tools
- aren't artists
Or both. That's why I encourage everyone to do the actual study of what's available with those tools, and get some intuition of what works and what doesn't, to realize that text to image (or text to video) isn't going to change anything. You won't be a "film director" until we have something that can pass for AGI, and probably even then. When trying to do it with text only, you'll end up writing and tweaking a weird essay, wasting a lot of time and not getting what you want. Content generation from text is always going to suck compared to combined compositing and ML techniques by someone with actual hard skill in arts. "Prompt engineering" is a gimmick.
That's not because those tools currently suck at understanding the natural language, although they do and will for years to come unless there will be several major breakthroughs in NLP, or the hardware will be a couple orders of magnitude more capable (as Google Parti demonstrates on their landing page), at which point it'll be much more transforming than just leaving artists without jobs. And it's not because the tools are limited by the dataset in what they can do (any concept not seen during training is guesswork, and it's very poor at it; for example SD sucks at dynamic posing). That's because text is just a poor medium to explain your intention and to control the result, even if you have a human on the other side. For still pictures, smart photobashing with style and concept transfer (finetuning the model on your own visual language) is where it's at, and for that you need actual understanding of visual arts - dynamic posing, anatomy, color theory, and generally artistic studies of how the world works. The AI tools will be developing in that direction.
Sure, a digital painter will replace a brush with something smarter, becoming more of a technical artist, but what's wrong with that? It already happened with 3D CG imagery 30-40 years ago, the doomsayers were also there, except we didn't have social media to amplify the panic. But the reality is it's always been an insanely complex and skill-demanding field. ML image generation is just another, higher order tool.
>There is an upper limit to how many images in still or moving form the world needs. We’re operating at that limit already - entry level jobs in the arts are often not paid or extremely poorly paid, signifying that demand for the jobs outstrips supply.
Currently, most visual entertainment is very much limited not by demand, but by the amount of work required. Japanese animation turned into a huge industry (and calcified into a more or less uniform visual language, compared to 90's) because it was cheaper to produce and dominated the world with tons of content, not because it's visually appealing more than other styles. Low resolution pseudo-pixelart dominates indie gaming because it's cheaper to produce with limited manpower, not because everyone buys the lo-fi/retro aesthetics. The ML-powered tools have a great potential to break these limits.
As an example, consider the particular niche which is believed to be disrupted immediately by Stable Diffusion - commission artists making custom character portraits for roleplaying sessions. SD can readily make passable fantasy portraits for this purpose (save for precision), but it needs a lot of work for that, and people still end up paying artists to do this work. This is not going to change for many years, at least. Moreover, the demand is also shifting to other styles, views, and poses with some thought put into it by a human, or potentially to animation/interactivity. The expectations are rising.
Visual entertainment is mostly driven by novelty.