Hacker News Reader

Comment by fab1an | Hacker News Reader

Amazed we haven't seen more DALL-E / Midjourney on HN. Probably the most astonishing new tech I've used since booting up a computer in the 90s.

Just generating images is barely doing the tools justice though - you can create entire mini movies with it, like SALT (a 70s sci-fi adventure happening on Twitter): https://twitter.com/SALT_VERSE/status/1536799731774537733

DALL-E's inpainting feature is incredibly powerful to generate very large scenes: https://twitter.com/fabianstelzer/status/1545752145273802752

Hard to believe that we're only beginning to scratch the surface here...

View on HN

I made a post on trying to do absurd-but-controlled food photography with DALL-E 2 (https://minimaxir.com/2022/07/food-photography-ai/ ) which did get upvotes on HN but apparently did not make the front page (so may have been flagged?)

Because if you spend more than 5 minutes with it you can determine numerous things about the "astonishing" images you have seen

1) the amount of human curation is huge. For every 1 good image shared there are dozen of utter crap not shared

2) Dalle fails in some very systemic ways that make it completely unsuitable for vast swathes of image generation (for instance the "N kittens problem" . Dalle is amazing at generating a picture of 1 kitten. Dalle is dreadful a generating a picture of 8 kittens and that is totally fundamental to how it works, not a bug that can be worked out with time.) Also basically anything that requires recognisable detail in the background, Dalle falls flat.

3) prompt parsing is simultaneously hit and miss as well as laughably primitive. This is the "without" problem. Ask for a picture without some feature and there is a good chance you will get that thing in the picture.

here's one i did recently: https://vimeo.com/724055394

a reading of dreamtigers by jorge luis borges

images created using @openaidalle. sequencing and morphing in #python with credit to András Jankovics morphing library [github.com/jankovicsandras/autoimagemorph]. featuring borderlands granular synth (artist template: @kingbritt), other desert cities delay by audio.damage , #rymdigare reverb, mixed in #kymaticaaum.

headphones recommended, awards eligible

(text here: thefloatinglibrary.com/2008/09/02/dreamtigers/ )

Because it’s a novelty abused for attention with clickbait popculture references. It’s not ready for primetime.

"photography is a novelty abused for attention, it's not ready for primetime and cannot replace a great painter"

and it was true until photography matured just like it is now with art generating AI. At least have a look at the early technology of your strawman argument. A better analogy would be the switch to digital photography it was really exciting because of the ease of use but years down the road nobody could use their early digital images for anything but stamps because of their atrocious quality.

Photography is an instrument the results of which are strongly dependent from the ability of the artist and technologist adopting it.

To the best of my understanding, DALL-E offers limited control and cannot be compared to a brush with paint, a photocamera, a virtual canvas for curves for illustration, a coding console.

Why? Because the weight of the user, its "importance", its "impact", is limited with that tool. (A commissioner is not an artist. A photographer may be.)

As HN member Moe wrote, «Bad analogies are like Vietnam».

Some people are already using generated images to replace stock art in low budget zines.

Reply on news.ycombinator.com