Technology

Honestly the more I tinker with ai image gen the more I get it. It's fun, but it's a different fun to doing it by hand.This is more tinkering and gambling, it def has a "one more turn" feel to it to see what the tool can come up with.

And the more I toy with it the more all barriers between whether this is art or not art break down, imo. There's a lot of work that can go into AI gen, the same way you can pick up a pen and doodle in the margins of your notebook, or you can learn ink techniques and make manga. You have the models (literally hundreds of them and then also merged models that combine 2 or more into 1), but also loras which skew the result in a certain way, you have to select the sampler (what creates the noise) and the sampling method (how much noise is added each step), etc.

Did I not make it, because the output comes from a computer? No, I did and I totally get why now that I've dabbled, even if this will ruffle some feathers. Does the producer not make the beat because he "only" sets up synthesizers in Ableton other people made and then plays chords other people found out before him?

It's not just typing in the prompt either, because the words you type either the positive or negative prompt have an influence, and part of it is also figuring out these words. You can also give them different strengths and you have to type the way the model expects to receive information. It's like a puzzle, though like I said there's randomness to it. A picture you feel looks cool (though I totally get if other people don't share the enthusiasm looking at it lol) could have taken a few hours to get just right. There's something personal to it, because it's my combination of prompts and parameters that spat this out.

The castle in the OP picture was made with a certain sampler. Here's two more:

The only difference between those three pictures was the sampler algorithm - everything else (the seed, the steps, the prompt, the resolution) was the exact same. This is just one way models can give wildly different results.

There's a lot that goes into it and look, I'm from a graphic design background and non-designers who made their own logo or used clipart they found on google to make one are nothing new. If something's good, it's good. You know it when you see it. Artists will use it as part of a process and then throw it into Photoshop or other software to bring the finishing touches. Others will just take whatever the AI spits out and say good enough. Corridor Crew used genAI that's available in a smartphone app to green screen themselves out without an actual green screen and even change themselves on camera in different ways, all from the raw video footage and a prompt. they made a short movie with it using the rest of their VFX skills; someone who knows what they're doing will use this as part of their process.

And beyond art and image gen there's the very undeniable tech aspect as you found out as well. If you want to get something specific out of Youtube but there's no solution, then just make your own with deepseek. If you want open source software but it's not working on your machine, deepseek it instead of going back to proprietary.

Lol you're the closest, and not that far off. Most people said 3-9 minutes. The answer is actually:

With proper GPU offloading (deepseek told me which launch arguments to use), pytorch and of course the GPU architecture probably plays a part in it. The model is actually meant to make 1024x1024 so that's what I generate at.

That particular picture was me trying different samplers to see the differences, the model's own limitations create a lot of artefacts but that's why I wanted to try Flux, and... it's scary:

This generates in around 40-60 seconds, I have TeaCache which cuts generation time in half but sometimes it takes a while for the model to parse the prompt with the TeaCache workflow, which is a bottleneck and I'm not sure why it's doing it. Flux also handles text but I'm not quite sure how to get it working.

Anyone who's scared of this tech... get acquainted with it intimately. It's not going away; this runs on the same computer that generated the picture in the OP and it's completely free.

Exercise caution and don't hesitate to search up the commands (both on google and in another chat window) before executing.

since this is a virtual environment I'm pretty safe. On VPSes I make sure to double-read the command before running anything, and do diagnostics first, like anything I copy from the internet. llms are very bad at asking questions before jumping to an answer so give it as much context as you can before anything, including OS.

I also try to make sure I can undo whatever it asks me to do if need be, which is actually easier to do when you have a single window with every instruction in it over a dozen tabs with some I haven't opened in over 45 minutes while troubleshooting a minor problem in a bigger problem lol.

I'm not going to spoil the answer yet but I just wrote it on another comment reply in this thread 🤐