this post was submitted on 25 Aug 2026
223 points (97.4% liked)
Technology
87550 readers
2985 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related news or articles.
- Be excellent to each other!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, this includes using AI responses and summaries. To ask if your bot can be added please contact a mod.
- Check for duplicates before posting, duplicates may be removed
- Accounts 7 days and younger will have their posts automatically removed.
Approved Bots
founded 3 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
I think this is a dangerous misconception.
I could feed locally run Minimax H3 a set of reference art, like the painted stills from this game, and it would exactly match the style. I could use start/end frames, motion masks, pose control, character sheets, voice and music samples, reference videos; I can get it to do whatever I want with extreme precision, and as long as I checked the output it would be indistinguishable from CGI.
The point I'm making is that a lot of people percieve diffusion as mashing a Stable Diffusion keyword prompt into a text box, and getting a sloppy result. Maybe thats still how shit corporate services like ChatGPT work.
But a LOT has changed over two years.
Theres no way people can definitively detect generative output anymore. We can IF its done lazily, but its quite possibly indistinguishable now, too.
You’re asking them to learn and understand rather than parrot… god speed.
Even two years ago, I remember someone telling me they saw a YouTube video that said you could always tell an AI generated image because their noise map would be the same in R, G, and B. And I tried to explain to this person that that depends entirely on how the image was generated, on what platform, by what model and method, what post-processing was done, etc.
And I was told I was ignoring their evidence. 🙄
Also... you could just add grain via post processing?
Its actually a reasonable technique, as it softens diffusion blocking artifacts the same way it softens video codec blocking artifacts.
The YouTuber sorta kinda has a point, depending on the context. Spammers tend to use the absolute laziest lowest common denomenator, so its quite possible theyre largely using the same model, with the same artifacts, and not bothering to edit output. So maybe they did have a common noise pattern? Though the way thats presented still sounds fishy to me.
I think the idea was that within an image the noise in red would match the noise in green and the noise in blue, because whatever system they were looking at picked a single noise map per image? Something like that. And yeah, it was reasonable because it was absolutely true of the laziest stuff out there (which is most AI stuff, which is why we use the word “slop”). But my counter-argument wasn’t that the method couldn’t work, it was that it couldn’t always work. Because image generation is done in thousands of different ways and post-processing exists.
The thing that I wish was better understood is that “slop” = “low effort.” Artists can make some pretty amazing stuff when AI is merely a part of their toolset and they still put effort into their work. But that means that not all AI is slop, and thus if you really want to detect AI, you can’t rely on it always being slop.
That is called luma noise, yes. Heh. Its nothing new.
Yes! I would wear this on my forehead on Lemmy, if I could.