NVIDIA is back with another #ML paper on their eDiff-I diffuser.
They seem to have cracked rendering text in generated images by using a combination of denoisers.
Stable Diffusion and DALL-E 2 struggle with text where eDiff-I succeeds (see comparison image.)
@don_kosak The photo on the bottom left begins to trigger a weird danger-related response from me 💀
@don_kosak Oh I'm sorry. I meant that as an interesting fact more than a complaint. It was intriguing
@don_kosak Thank you, that was kind.
'Uncanny valley' — that's what I was meaning.
@golemwire A lot of AI art is still firmly in the "uncanny valley". I wince every time I get a generation with extra arms, legs or fingers. 😅
I wasn't sure how to interpret your first comment, but I knew AI images have a lot of different responses.
I CW'ed the image in my follow up post to be on the safe side.
I'm new here, so I'm still getting used to the platform.