There's some people who are just churning away StableDiffusion for days and finding the 'perfect' one-in-a-million image, but there have been drastic changes in the available tooling, too, and even moderate use of those better tools can save you a ton of time and be easier than filtering through a pile of rejects.
Inpainting lets you reroll (with different prompting) just the particular portion of an image that has flaws or doesn't match your goals, ControlNet can intake data in a wide variety of formats to provide information like pose or depthmap or a user's 'scribble', LoRA can find the vector representing parts of a group of image that may not have easy human terms, so on. Some things, like regional prompting or psuedo-panographs made from multiple initial images then stitch/merged into a seamless whole, can produce images that are just not possible to do with simple prompt-only AI-gen.
These advances aren't able to solve every limitation of the tool: AI doesn't have, and maybe can't have, taste. It's not that it can't produce reasonable hands, but that it can't tell you that you should care that this malformed hand matters; it can save a real artist from needing to break their wrists drawing rosettes, but can't tell when to stop; it will quite happily shade in some of the most eye-searing colors that the greatest fan of sparkledogs would flinch at. (Similarly, it's hard to get LLMs to avoid purple prose.)
The more serious issue's the social and economic effects. Even assuming AIgen never gets to the point where it can reliably produce some coherent Great Art (or even just a short-form comic), it's absolutely enough to break a lot of demand for at least some meme-tier or 10-dollar-sketch grade work -- and even the Creative-Common/public-domain-based models like Mitsua can do that. And a lot of artists either pay their bills or broke into bigger-value commissions thanks to that sorta work.
Socially, both the furry fandom specifically and the general online art sphere doesn't really have a good way to handle a rando putting up 10k pieces a day, every day, in every community. And if you could just ban that jerk, you can't handle a hundred people putting up hundred. Not in discoverability, not in moderation capability, not in simple hard drive space.
((I don't buy frustrations about the environmental costs, especially on runtime side. SD1.4 took a lot less energy to train than most people think, and you can render individual images for thousandths of a KWh. And I'm skeptical that people would accept the revised models that took an order of magnitude less to train; there's model equivalents trained from scratch for less than $50k, and that includes the cost of renting equipment.))
LLMs (ai text) are... rough; as bad as artists see the flaws in even good AIgen visual media, even better LLMs tend to be pretty bad about not actually giving what you prompted for. I wish they were at the necessary point where they could at least be useful, but they don't handle complex scenarios without a ton of micromanagement or dropping in random new features to a scene -- Lowd's description of it as 'autistic' isn't the best way to describe it, but there's very much a tendency for it to infodump randomly at you. There are some writers that use it to supplement their original work (afaik, none in the furry fandom... though given the rate people who are out get blackballed, how would you know?), like Leanne Leeds, but they tend to either only use it to augment details where they're more for setting and pacing than plot, or shipping tons of cruft, or both.
There's some people who are just churning away StableDiffusion for days and finding the 'perfect' one-in-a-million image, but there have been drastic changes in the available tooling, too, and even moderate use of those better tools can save you a ton of time and be easier than filtering through a pile of rejects.
Inpainting lets you reroll (with different prompting) just the particular portion of an image that has flaws or doesn't match your goals, ControlNet can intake data in a wide variety of formats to provide information like pose or depthmap or a user's 'scribble', LoRA can find the vector representing parts of a group of image that may not have easy human terms, so on. Some things, like regional prompting or psuedo-panographs made from multiple initial images then stitch/merged into a seamless whole, can produce images that are just not possible to do with simple prompt-only AI-gen.
These advances aren't able to solve every limitation of the tool: AI doesn't have, and maybe can't have, taste. It's not that it can't produce reasonable hands, but that it can't tell you that you should care that this malformed hand matters; it can save a real artist from needing to break their wrists drawing rosettes, but can't tell when to stop; it will quite happily shade in some of the most eye-searing colors that the greatest fan of sparkledogs would flinch at. (Similarly, it's hard to get LLMs to avoid purple prose.)
The more serious issue's the social and economic effects. Even assuming AIgen never gets to the point where it can reliably produce some coherent Great Art (or even just a short-form comic), it's absolutely enough to break a lot of demand for at least some meme-tier or 10-dollar-sketch grade work -- and even the Creative-Common/public-domain-based models like Mitsua can do that. And a lot of artists either pay their bills or broke into bigger-value commissions thanks to that sorta work.
Socially, both the furry fandom specifically and the general online art sphere doesn't really have a good way to handle a rando putting up 10k pieces a day, every day, in every community. And if you could just ban that jerk, you can't handle a hundred people putting up hundred. Not in discoverability, not in moderation capability, not in simple hard drive space.
((I don't buy frustrations about the environmental costs, especially on runtime side. SD1.4 took a lot less energy to train than most people think, and you can render individual images for thousandths of a KWh. And I'm skeptical that people would accept the revised models that took an order of magnitude less to train; there's model equivalents trained from scratch for less than $50k, and that includes the cost of renting equipment.))
LLMs (ai text) are... rough; as bad as artists see the flaws in even good AIgen visual media, even better LLMs tend to be pretty bad about not actually giving what you prompted for. I wish they were at the necessary point where they could at least be useful, but they don't handle complex scenarios without a ton of micromanagement or dropping in random new features to a scene -- Lowd's description of it as 'autistic' isn't the best way to describe it, but there's very much a tendency for it to infodump randomly at you. There are some writers that use it to supplement their original work (afaik, none in the furry fandom... though given the rate people who are out get blackballed, how would you know?), like Leanne Leeds, but they tend to either only use it to augment details where they're more for setting and pacing than plot, or shipping tons of cruft, or both.