[metalcrow]

The author claims that this kind of images shouldn't be in the training data, and agree or disagree with that, I'm unsure how much removing it would actually prevent such images from being generated. AI can certainly cobble disparate concepts together quite well, it seems unlikely violent and visceral images couldn't be regenerated from other non-violent content.

[km3r]

I think it speaks to the unfamiliarity the author has with the workings of AI. A misunderstanding of the latent space and how it can generate bizzare images when it has little to go off of or inverse negative directions.

[nozzlegear]

AI can barely figure out how to make a cartoon pelican ride a bicycle.

[bobsmooth]

Generating SVG code and generating an image are two different things.

[nozzlegear]

What would the LLM generate more accurately: an svg of a pelican on a bike, or an svg of a gory, dead woman?

The medium is superfluous.

[BoorishBears]

> The medium is superfluous.

No it's not. Generating an SVG is asking the model to write text that can be rendered as an image.

Generating an image directly skips the intermediate step and directly outputs an image: so the accuracy is in a completely different league. The models people ask for SVGs typically cannot do this, only specially post-trained variants can.

(An LLM will do the svg of a pelican on a bike much more accurately btw.)

[fragmede]

AI does fine at that. LLMs have problems generating SVGs of that, but that's kind of an (intentionally) particularly obtuse test.