News for hackers. Stuff that matters.
Re:Ten Years of Image Synthesis by Keyframe on Tuesday November 29, 2022 @04:02PM (#33793250)
What's next though? To paraphrase Hemingway - It all happened gradually and then suddenly.
Re:Ten Years of Image Synthesis by khazhoux on Tuesday November 29, 2022 @04:38PM (#33793691)
My mind is blown away by Midjourney v4, but the open problems are still plentiful (even before getting into video/motion, 3d etc).

After a few dozen prompts, you quickly learn how the image generation is simultaneously amazing and dumb. The dumb part is the next 10 years of improvements.

This is like Google 1999.

Re:Ten Years of Image Synthesis by ajmurmann on Tuesday November 29, 2022 @06:58PM (#33795095)
Just getting the number of fingers and to a lesser degree arms and legs correct would solve 80% of the issues I encounter
Re:Ten Years of Image Synthesis by khazhoux on Tuesday November 29, 2022 @07:48PM (#33795465)
Right. And imagine being able to refine an image with the instruction to move the left arm down a little, and lift the chin up just a tad. Today that's impossible.
Re:Ten Years of Image Synthesis by fjkdlsjflkds on Wednesday November 30, 2022 @12:57AM (#33797468)
Re:Ten Years of Image Synthesis by khazhoux on Wednesday November 30, 2022 @04:41AM (#33798723)
Very cool, thanks.
Re:Ten Years of Image Synthesis by theemathas on Wednesday November 30, 2022 @03:15AM (#33798245)
Seems really close to reality https://prompt-to-prompt.github.io/
Re:Ten Years of Image Synthesis by tintor on Tuesday November 29, 2022 @08:40PM (#33795854)
Legs, arms, and heads correctly oriented too.
Re:Ten Years of Image Synthesis by theemathas on Wednesday November 30, 2022 @03:10AM (#33798206)
Re:Ten Years of Image Synthesis by khazhoux on Wednesday November 30, 2022 @01:02PM (#33804771)
Some of those hands are ok/good, but even the second image in that collection has messed-up hands.
Re:Ten Years of Image Synthesis by ajmurmann on Friday December 02, 2022 @09:24AM (#33830745)
I have generated some images with sound hands. However, I've had even more come out with defective hands.
Re:Ten Years of Image Synthesis by planetsprite on Tuesday November 29, 2022 @04:08PM (#33793343)
In the arts, music, graphic design, animation, 3D models, and eventually realistic looking video seem to be the clear next steps.

The key to AI art is that while the universe is infinitely complex, the visual form of most things humans can recognize follow a collection of natural patterns. Anything humans can do can be computed, the only issue is being able to set up the scope for the patterns we seek to replicate. For games like chess, a bit easier, for images that represent label-able things, harder but still manageable. The main problem of making things beyond this is giving a valid scope to the whole problem: what patterns, precisely, do we intend to replicate?