News for hackers. Stuff that matters.
Re:Illustrating Gutenberg library using Stable Diffusion by temp_account_32 on Sunday September 04, 2022 @12:04PM (#32714446)
I feel like the next killer feature of these models will be when they can retain contextual visuals between different scenes being generated

As in, the same mouse in four different pictures in different situations, not four different mice styled in the same way over four pictures

Re:Illustrating Gutenberg library using Stable Diffusion by karencarits on Sunday September 04, 2022 @12:14PM (#32714553)
Textual inversion for stable diffusion, which is related, was discussed a few days ago: https://news.ycombinator.com/item?id=32643564
Re:Illustrating Gutenberg library using Stable Diffusion by temp_account_32 on Sunday September 04, 2022 @12:30PM (#32714748)
This is nice but still has some huge limitations, e.g. you need 3-5 images of something to train it. It becomes a chicken and egg problem if I start by generating something, since I cannot create 5 different images of it to train the textual inversion. It is also intensive and slow to train on each object.

Also I think we're running into limitations of using a single prompt to describe things, as it's often hard to separate the 'what' is in the scene and the 'how' to draw the scene, with many prompts looking like a giant word salad trying to tap in the internal secret language of the model

Re:Illustrating Gutenberg library using Stable Diffusion by hirundo on Sunday September 04, 2022 @02:45PM (#32716334)
The same tool can also be used to evolve an image by generating variations and selecting a few to reproduce. You could breed and cross the results at will. You could use multiple preference models for the fitness test and evolve them afk.

An "Evolve" button could become a standard feature on content editors across media.

Re:Illustrating Gutenberg library using Stable Diffusion by dr_dshiv on Sunday September 04, 2022 @12:49PM (#32714988)
Yeah, this seems to be the key challenge right now. Some how an image and its elements might be used as context to condition the subsequent generations.
Re:Illustrating Gutenberg library using Stable Diffusion by dr_dshiv on Sunday September 04, 2022 @03:25PM (#32716743)
Oh, looks like Google solved this last week with dream booth: https://arxiv.org/pdf/2208.12242.pdf