Be advised: this thread is not for serious in-depth discussion of weighty topics (we have a link for that), this thread is not for anything Culture War related. This thread is for Fun. You got jokes? Share 'em. You got silly questions? Ask 'em.
- 217
- 1
Jump in the discussion.
No email address required.
Notes -
Yeah, it's starting to show up everywhere that the specifics of the output art don't matter.
And it may not be just the cost driving that: working with or around artists has a lot of other failure states on top of it. However bad the intellectual property concerns for AI artwork are, it's not as likely to literally trace Devil May Cry or Touhou fanart unless you put a lot of effort into making it. It's a lot less likely to show up with the assets three weeks late. And barring
mechahitlerGrok, you don't have to worry about your imagegen tools ending up the high-profile example of the day on Twitter.What really bugs me is where people use AI art despite having actual assets that would be more representative and credible. Minecraft modpacks and content creators have heavily bifurbicated between the users and the antis, like everywhere else, but there's further split between people who use it as if it's the only images they have, despite screenshots being more useful and relevant. This might be slop on top of a well-executed traditional trailer video, but then you see stuff where there's just slop the whole way through, as if Ben Garrison was in charge of their marketing.
Eh... I dunno. The gooner market seems like it should be the most maximally able to do it (and, bluntly, having the lowest standards), but it really hasn't gotten there, there. Even if the art can 'fool' you, or be what would be technically impressive for a human to make, that doesn't make it interesting or good in the way that an MSPaint scribble can be. This (artist: pantheggon. cw: f/f, mind control) is the sort of thing that an AI-gen model could easily do, in terms of artwork quality, but it's also notably not the sort of thing anyone's actually done with AI-gen. Even where you let the machine build direction for you, and , .
It may be we end up with more of a muddle, where the higher-value the effort, there more buyers can justify the cost of human intervention, and that ends up a sliding scale of completely AI 'junk' content, intermediate works where there's a lot of human planning and the AI is doing the busywork, and then at the high end stuff that's computer-mediated only in the same sense CGI is.
That's interesting: the really harsh tell for me is the pacing. You can 'beat' a model into playing a little better toward a specific tone or form of speech pretty easily with an exemplar file or some pretty basic OpenCode setups. With the correct subagent layering, you can kinda get it to switch tones or character voices from one chapter to another. It's really hard to get it to stop treating each subsection as a crescendo of its own, and that gets my hackles up.
Some of that seems downstream of the tools being very incomplete. I'm still finding the most effective option is OpenCode and a handful of custom skills, and it's still not good there so much as indicative that mainstream processes suck, but it points.
Of course, the default-version stuff that sucks is already selling enough to be a concern.
(context, probably?)
I'm a little curious, here. I didn't push those sessions very hard on prose quality, since I don't really want to replace my tone with the LLM's form, but a lot of the advice on plot notes, character psychology, and specific actions seemed pretty specific.
Like, I'll admit AO3 has things better left unimagined, but 'your crossdressing female main character hasn't considered all the aftermath of using a men's urinal, then is surprised by a gay guy being uncomfortably close to her' seems like it's not overrepresented in the training data, and it's an important enough scene that I did try throwing a version with intentional errors and the model caught them. The Diogenes/Stoic thing is more mainstream knowledge, but actually drawing the thread through a character who's pretending at self-abnegation and instead undermine his ability to communicate. Or stuff like figuring out whether a bench press is being described correctly, or how well I've extrapolated how different species might adapt to swimming.
They're not always right, especially when 'right' is defined as 'the story I'm trying to tell', but it's a far cry from just throwing 'tighten your prose', 'X character needs more depth' and a dozen em-dashes out.
I agree that they're very prone to positivity and agreement bias ('glazing') and to the standard AI ticks, but if I'm not intending to use the prose from that discussion, that seems pretty easy to overcome. It'd definitely be better if they didn't, and the few times I've gotten genuinely critical analysis has been a lot more rewarding.
The more complete claudemd or OpenCode versions are a lot stronger, especially as the input work increases.
Yeah, they're a lot better as original-level beta readers than final ones. Doubly so for genre works, but even general-audience stuff they're still more capable of saying stuff about what an average (if very broadly-read) human would think, rather than your actual readers.
Oof, very much an increasing thing in a lot of circles. Writing, RPGs, video game mod artwork, code discussion. And there's a lot of downstream mess: a number of environments are adding AI-assisted tags, and then you get people who bulk-block everything with it whether they're using it for a few minor pieces or just doing an edit pass, and then you get people who don't use the tags even where their entire project is clearly AI-gen, and then we end up in a situation where users who are legitimately trying are invisible to anti-AI people and the spammers are all that get through the gaps.
And then the loudest advocates tend to be Asmongold or Shad-likes, regardless of their volume, because the obnoxious ones get the most publicity and repetition
More options
Context Copy link