Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[theDragn]

Learning to use AI art tools...

That first one looks a lot like Apex sprites. While I have no idea if you used those sprites as training data or not, please be aware that I specifically do not give permission to use them in that manner.

[Lukas04]

I am not completly against the usage of AI if you just use it as a tool to draw inspiration from, but by the looks of it you used a lot of modded content for your model, and if thats true, not mentioning which those have been trained on feels a bit wrong, even more so in case the artists havent been asked for consent from.

[Dazs]

I am not completly against the usage of AI if you just use it as a tool to draw inspiration from, but by the looks of it you used a lot of modded content for your model, and if thats true, not mentioning which those have been trained on feels a bit wrong, even more so in case the artists havent been asked for consent from.

Being a fine purveyor of Spiral Arms I am very familiar with the free to use content here and I can see the majority of the AI sprites sources are from this thread. Xenoargh is the OG curator of free to use and I highly doubt any mod released sprites were used.

[Lukas04]

I am not completly against the usage of AI if you just use it as a tool to draw inspiration from, but by the looks of it you used a lot of modded content for your model, and if thats true, not mentioning which those have been trained on feels a bit wrong, even more so in case the artists havent been asked for consent from.

Being a fine purveyor of Spiral Arms I am very familiar with the free to use content here and I can see the majority of the AI sprites sources are from this thread. Xenoargh is the OG curator of free to use and I highly doubt any mod released sprites were used.

I do hope thats true, but at my own glance, the Generated ships do not look like they were trained on data from this thread. Most of this Thread uses Kitbashed Sprites, or Sprites following the original starsector style quite closely. Despite that, the plating on those generated ships very much doesnt look like its been trained on just that. They replicate more so the sprites of people attempting their first non-kitbash sprites, which often have more rough lines and not the best transitions inbetween parts.

Thats not to say that AI would perfectly recreate them of course, but ive seen some AI trained on more Vanilla adjacent data and they tend to create them quite differently, though in the end that could also come from the training process.

But rather than speculation it would probably just be better to hear the response from the author.

[BigBeans]

My Star Wars ships I made a long time ago that I no longer want to maintain. Maybe someone else can find use for them.

[JAL28]

Suppose there's a few things I can sacrifice to the maw of endless hunger:

Spoiler

Yellow Warthog. That's it. It's a yellow Warthog.

Unimplemented carrier with probably the most bay/volume in the entire modiverse. Was originally going to ES but got rejected so throwing it to spiral arms instead.

[close]

[xenoargh]

Those Star Wars ships are lovely. I'll get new submissions into the official archives... er... soon?

Speaking to the AI-art stuff: Rest assured, I'm not secretly training a model on mod work, lol. My goal was simply to figure out how to make stuff that looked a bit like a SS ship very quickly. I've discussed this in this thread, but I have lacked time to provide a lot more information; I'm currently working on an experimental tool to "talk to" Large Language Models in more effective ways IRL.

Looking at the older stuff at this point makes me kind of laugh, honestly, because the stuff I've been able to get out of Easy Diffusion at this point is a lot more advanced and polished-feeling, and the quality of the img2img input image doesn't have to be all that great.

Take a source image used as a design concept, such as this:

AI Image Source Note: these images were all made with img2img prompts and Stablity AI's Stable Diffusion, using the default model. Text prompts were probably something like, "spaceship, panel lines, greebles, small lights"

Add in properly-designed prompts, use a model that produces fairly good results, and et voila:

If the results feel a little too "wild" (i.e., not close enough to my design concept) I can always tighten up the amount of random noise and get closer to the original concept:

Rinse and repeat until you're happy with the results, or just kitbash things to get all the major details you wanted.

So, as I hope you can see, there's no need to secretly use modders' work to train AIs with or anything remotely nefarious like that. I don't even need to use a complex design to get reasonably-close results. This is just what the future looks like, where humans provide the big ideas about design and AIs can be trusted to fill in a lot of smaller details.

As usual, if you like any of that stuff, above, feel free to use it like anything else I've put into the thread, including the image I used to create these works with, if you want to explore this technology further.

[theDragn]

Speaking to the AI-art stuff: Rest assured, I'm not secretly training a model on mod work, lol.
[...]
So, as I hope you can see, there's no need to secretly use modders' work to train AIs with or anything remotely nefarious like that.

I'll be honest, I'm not very assured, because you just said "nuh uh, I wouldn't do that", but didn't show the inputs you used to generate the initial ships. Regardless, I'm not looking to have a fight about whether you actually did or did not- if you did, don't do it again; if you didn't, then there's no problem.

As far as this being what the future looks like... well, it looks like crap. None of those ships make any sense; it's just a bunch of vaguely-plausible details that fall apart when you take a closer look. You could kitbash something similar and better out of vanilla derelict sprites.

[xenoargh]

Tell you what: I'll just let the entire cat out of the bag, so you (and everybody else here) can play with the AI stuff and replicate my results with exactly the same software and information.

I didn't want to provide a lot of detail in public about this yet, because I'm still very ignorant of how best to prompt-engineer for perfect results, and frankly, I feel my understanding of this subject is laughably primitive. There are probably guides online by now that explain this stuff so much better than I can. But hey, this should reassure you. I love riding the bleeding edge and I'm having fun, that's all!

To use this, you'll need a CUDA-capable GPU, i.e, nVidia with enough VRAM (RTX Series 10+, technically; I'm using a 3080 TI).

First, install Python and Easy Diffusion.

Once you've installed it, run it; it'll open a web browser page. The first time, you may need to tell your OS / AV it's cool, because it does Scary Python Things and talks to the Internet and installs sub-components, etc.

The settings I'm using, other than the text prompt part, are:



The text prompt used is: "post-apocalyptic spaceship, [aerial view:2], panel lines, greebles, hyperdetailed, small lights, small details, dirt, grunge, damage" sans quotes.

If you're trying to replicate my results exactly, you'll notice that dread_V2_dread_V2 isn't available on the Model dropdown tab of the UI. I don't think you need this model to get good results, but here's where you can find that, if you're interested. I'm using that mainly for other things, but it works fine for this purpose.

The earlier stuff I've shown was using Stability AI's default model 1.4 or NightCafe's custom model.

[EDIT]
Oh, and!

Somebody, if they're trying this out for themselves with these instructions and their own art, might notice that all my starter images have weird noisy texture overlays. I've found that that (and starting with an image that's about as large as the image you intend to make, if not larger) makes much, much better images. Why? Because of how the AI's initial pixel-sampling methods work, basically. I just use a "tech" texture I have sitting around, but I'm pretty sure most forms of noisy grayscales used w/ Multiply (and some Curves to pull up white points afterwards) will work.
[/EDIT]

So there you go. No super-secret AI model stuff is required. People have been doing this kind of stuff for quite a while now; I'm just providing some basics, and have probably even given some bad advice out of sheer ignorance; but these settings have worked for me, for the kinds of things I'm doing.

Another example. Let's try and make something vaguely Midline-ish. Is that "Fadur" for 'fake Kadur' or "Midfake"? IDK, lol.

Starter image:


Some results:

[Harmful Mechanic]

You could also run https://github.com/eballai/NoAI or https://glaze.cs.uchicago.edu/ on your images, if you're worried about having them scraped for use in a model like this.

[xenoargh]

Right. Like, just so it's utterly clear, though, and at the risk of repeating myself again.... what I'm doing didn't involve any models that aren't already public and freely available.

I wrote a fairly lengthy bit earlier about how much drudge-work is actually involved in using Dreambooth, but frankly, it's pretty technical and I didn't think it would be as convincing to non-techies as simply showing exactly how to do this themselves. People think this AI stuff is magic; it really isn't. Training an AI on all the imagery just in this thread would involve writing hundreds of very carefully-crafted JSON descriptors, then using a pair of A6000's to process a model... and IDK about you, but I don't have time for that, just to make some SS sprites, lol.

[theDragn]

You could also run https://github.com/eballai/NoAI or https://glaze.cs.uchicago.edu/ on your images, if you're worried about having them scraped for use in a model like this.

Oh, this is very useful. Thank you!

[BigBeans]

More old stuff I don't use.

[xenoargh]

Those are great!

Meanwhile, took a shot at Low Tech last night. This has been the biggest challenge to get an AI to interpret, in terms of intent vs. results. Low Tech's all the things that AIs tend to hate- non-stereotypical rocket forms, complex, almost biological-feeling greebles, etc. So I've really struggled with prompting and base images. I'll provide some example prompts / base images when I feel more confident that others can get good results.

For now, these are a few samples, including a couple of things I put turret positions and a few extra details on, just to show the transition to a final work.

AI Image Source Note: these images were all made with img2img prompts and Stablity AI's Stable Diffusion, using the model known as Dread_V2. Text prompts were probably something like, "spaceship, ((aerial view)), panel lines, greebles, (((detailed and intricate))), rocket engines, jet engines, small details, dirt, hydraulics, hatches, struts, grunge, rust, oil stains, post-apocalyptic, low-tech, steampunk, dieselpunk"

[SirHartley]

I don't think this whole AI talk should be in this thread, it feels like something that should be a separate one.