Print

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

[pixel_rice_bowl]

This is a collection of 109 portraits that tries--as best as I could--to stay close to the vanilla lore and art style. The quality varies such that I could probably pare this down to 70 or 80 portraits, but here I figure it's better to include more than less.

The collection took about 40-50 hours to make. It was created using a combination of Stable Diffusion plus manual cleanup. I generated around 4,000 portraits across a couple dozen prompts using varying sampling models and settings and then painted and edited them in Krita and Aseprite. Since Stable Diffusion often creates portraits with the anatomical equivalent of severe dmods, most of my time was spent drawing over or editing results from decent enough bases as well as adding or fixing portrait and rim lighting and errant pixels. Removing outliers, I want to say the average portrait took about 30 minutes to an hour.

Edit: Here's a picture of the full sprite sheet:



I'm releasing these to be used by modders for any packs or mods. You can find a collection of all portraits as individual files here:
here (edit: or as a zip here).

Someone is welcome to put them into pack and post them to the forums on my behalf if so moved. You can use, edit, or otherwise modify these in anyway you like. No attribution or credit is needed.

[Thaago]

Welcome to the forum!

Wow, these are pretty great. Its interesting to hear the amount of post processing needed on autogenerated portraits to make them good. Did you have to feed the program the existing portraits or anything as training data, or was it just text prompts?

[Kos135]

Looks good! Nice to see a portrait pack that isn't apart from the vanilla style, or even worse, anime.

[pixel_rice_bowl]

Welcome to the forum!

Wow, these are pretty great. Its interesting to hear the amount of post processing needed on autogenerated portraits to make them good. Did you have to feed the program the existing portraits or anything as training data, or was it just text prompts?

Thanks for the welcome! Starsector is one of my favorite games of all time and I've lurked here for a while.

It was a combination of text prompts and image feeding. The process fed into itself--once I found some good prompts, I used the resulting images to create new and similar images. Sometimes I used existing portraits, but with the way img2img works, it requires some pretty careful slider control to get something useful. (Last I read, training the models on specific image sets requires an industrial grade GPU--something closer to an A100 than a 3090--which isn't something I have access to.)

And, yeah, it usually requires a good amount of work to make a result look decent. I have a few years of pixel art and 3d rendering experience, which made the process smoother, but it's clear even with a lot of prompt magic that the vanilla portrait style won't be replaced by an AI. Many of the portraits are close in some regard but they don't have the same faction feel as vanilla and the overall quality is sometimes inconsistent (as I've followed reaction to the portraits elsewhere, it's clear I made a few basic errors on at least a few of them that a trained artist wouldn't have). These deficiencies are one reason I decided not to make them into an actual mod--it needs more quality control.

I'm sharing this set primarily to show others how much labor can be saved by generating concepts or workable bases. I hope skilled artists make use of these new tools to create some killer portrait packs.

[cytokine]

Wow, these are incredible!

[CrashToDesktop]

Well done! I do appreciate having vanilla-style portraits, I think I'll be adding many of these to my runs from here on out.

[Rusty Edge]

X-8, Y-5 has the haunted look of a Vangaurd crewman who has seen a battle or two.
X-8, Y-9 looks like an older Bobby Draper.
X-5, Y-9 A proper prince charming.
X-11, Y-5 I don't know why, I just love this guy.

I'm recognizing quite a few of these faces. Definitely an impressive collection, quite a lot of personality and possible stories here.

[TheLemu]

Finally something to replace those pesky anime/meme portraits mod authors push down everyone's throat lmao
Thanks for compiling these

[Brainwright]

It's a good demonstration, and I can see the judgement calls you made to improve various individual portraits.

However, I wouldn't use these.  The first thing that struck me when I saw them as a whole is that they look the same as the vanilla portrait set.  Same general shape and the same color scheme.

When I want to add a portrait pack, it's to expand the general range of shapes and colors to generate unique impressions.  Viewed individually, the AI generated images might seem to have variety, but it's very shallow.  You cease to be able to differentiate between individual portraits if you use them in-game.

That's been my impression of AI-generated art so far.

[cytokine]

The first thing that struck me when I saw them as a whole is that they look the same as the vanilla portrait set.  Same general shape and the same color scheme. When I want to add a portrait pack, it's to expand the general range of shapes and colors to generate unique impressions.  Viewed individually, the AI generated images might seem to have variety, but it's very shallow.  You cease to be able to differentiate between individual portraits if you use them in-game.

Ideally I'd want a distinctive look for my own officers, but for other NPCs, what the game needs less of is the 4th wall-breaking moments where you run into a clone of say, your own administrator. For me, that was the main downside of playing with vanilla portraits only. So you need more portraits to minimize duplicates. But if you do run into a duplicate, some blandness will cover it up, while distinctness makes it blatant: you would remember seeing exactly that face somewhere else, just because it was memorable. So quantity > generic > distinct, IMO. Against that kind of immersion-breaking, at least.

Not saying this portrait pack is bland. They just fit right in.

[Thaago]

I took another look and one thing I noticed is that I think the AI has a bit of difficulty in the size of helmets/hats: some of them are crazy big!

[Brainwright]

Not saying this portrait pack is bland. They just fit right in.

Not saying it's bland, it's that if you can keep most of these portraits in your memory, they start to blend together.  Your response to them is all the same.  It's the sort of thing a magician would use in a trick.

[cytokine]

It's the sort of thing a magician would use in a trick.

Well... perceiving the in-game world as vast and as real and as full of people is an illusion, and creating immersion means making that illusion believable. That's the goal. It's like, tricks all the way down, dude! Anyhow, "different strokes" and all that...

[Great Wound]

Someone is welcome to put them into pack and post them to the forums on my behalf if so moved.

Done!

[cytokine]

Done!

Oops, no faction file for the luddic path? I guess it would be identical to the one for the luddic church?

Since Stable Diffusion often creates portraits with the anatomical equivalent of severe dmods, most of my time was spent drawing over or editing results from decent enough bases

What are you talking about? I had a go at making starsector portraits with stable diffusion about a month ago, and those turned out perfect. Someone should make another portrait pack...

Spoiler

[close]