Brand new LORA training to replace the full model trained from ages ago. Works very well with multiple different models.
Use the keyword starsector to affect the images during generation all the examples were done with super simple prompts nothing more than "starsector photo of a ________"
Description
FAQ
Comments (23)
This seems great for creating portraits for ttrpg sci-fi games like Traveller. Would you be willing to do a Stellaris portrait model? That would be amazing for mods.
Oh hoho, yes indeed it would =)
Looked into it a little bit, all the Stellaris stuff is split up into pieces since they are all animated. If you have any ideas or suggestions on where to find decent static image portraits to make a dataset out of I wouldn't mind giving it a shot.
@severianvoid Hmm, not really, you'd have to convert them into static jpgs. There is a way, but I forget how.
@severianvoid I cannot link directly (have already submitted the link to the whitelist) - but since this is the official paradox Wiki - I'm rel. sure Civitai Mods won't be mad about this: stellaris DOT paradoxwikis DOT com FORWARDSLASH portrait_modding . It's not malicious or spam. You will find all the necessary resources there. From a first glance, the game will accept static portraits - in the *.dds format it seems. There should be no problem in finding tools that can convert into this container. If there is a way to compile the portraits into a static library (or to assemble usable ControlNet templates =), it's the best resource I could come up with. Good luck =)
@xpgx1 I am not sure if there is enough of each species to get good diversity compared to the variation available in the game itself. Only 12-18 images for each different species, but its at least a possible starting point compared to trying to figure out how to get images from all the various segmented out part images from the base game. I will see what kind of results come out.
@Taco360 @xpgx1 well I got results I was pretty happy with so new model published, hopefully it does what you wanted. Definitely had a hard time with a few aspects of it and its the first time I tried to put multiple trained concepts into a single training file but it works pretty well I think.
good work king.
That's a really great Lora, thank you so very much! Even though I already used AI to create Star Sector portraits for myself, it is always good to have a dedicated tool!
Any update for XL?
Been working on and off on a SDXL version for about two months now, huge number of different training runs so far and the results have not been that good though with a lot of iteration and dataset manipulation its getting better. SDXL has proven to be pretty troublesome to train compared to 1.5 and a lot of that just comes down to up-scaling 128x128 images to 1024x1024 for training. But also training on a lot more keywords to try and allow specifying individual faction styles which has greatly increased the complexity.
I should hopefully have a few people testing the best training candidate soon though. And then depending on how it goes post a release.
I also thought about it, but honestly... what's the point? If you're going to use it in your game, you'll downscale them anyway. That's why I don't see any point in SDXL one. The only reason to do such a version is because SDXL is better with certain concepts, but you can train these with a lora.
I have finally gotten results from training that I am exceptionally happy with, so I think barring any issues with testers there should be no more training. I am now in the middle of the task of narrowing down from the 5 best training runs, which of the checkpoints saved out at different steps performs the best to be the final candidate for uploading.
@Cheru I agree, by and large if you are modding the game you will be downscaling to 128x128 anyways so in that regard the sd1.5 version is just fine. However there are a couple of reasons to train a new version mostly just better tools and more knowledge. First thing I am very happy with is I was able to actually train the model to produce both high quality 1024x1024 images as well as have it generate a pixel style version of nearly the exact same image by using the terms "high quality" and "pixel art" in the prompt while keeping the other generation parameters the same.
The next thing I am very happy about is having additional keywords like the factions names, spacesuits, helmets, so a lot of new keywords that function in the newly trained model examples are things like "Hegemony, Luddic, Sindrian, Spacesuit, Helmet, Mercenary, Civilian" I don't remember the exact list off the top of my head, but it should really help quite a bit and I can't wait to finally post the final thing with all the various examples of what can be done.
@severianvoid glad to hear that. As I said, new concepts(+new skills in training) are the reasons for SDXL upgrade. What model you used as a basis for the training? Pony, basic SDXL 1.0, or something else?
I will soon unleash my 1,5k portraits on your gallery here and one of the most difficult tasks was to create a Hegemony set. At first I forgot their color palette and made them wearing green armour and uniform, then I actually checked and remade them in grey with orange trim, didn't like the result and decided to go back to green theme. In the end, I ended up with a mix of green minor NPCs and grey/orange major faction NPCs for the Hegemony. Will be interesting to see what this new lora can do in that regard xD
@Cheru SDXL1.0 is what I have been using as the base. The ability to make uniforms that match has been really good so far in my testing, using just a base term like "hegemony" will give you the standard tan, black and orange stripes. But you can still add colors like "blue hegemony" and it will change the colors around but keep the overall style of the uniform (shoulder pads, stripes across the chest, things like that). Really hoping that people can make use of the better training and more keywords.
I have been testing it in a number of models, and it works pretty well in most, fantastically in some, but seems to really struggle with a lot of the anime-ish ones, which I think are a lot of the pony merged ones? not sure I haven't looked at many other models outside of base SDXL so far. It may need a dedicated training run for some of those others. If you have any good model suggestions to test with let me know, I haven't been following a lot of the new models so well so I'm still trying to catch up on things like Lightning, Turbo, Pony and Flux.
So far besides the base SDXL model JuggernaughtXL has performed incredibly well. AnimagineXL as an example though struggled quite a bit.
@severianvoid sounds very flexible and intriguing! I'm just a bit terrified right now, if that's really so good, I might redo my collection xD I tend to get... carried away, so to speak.
In terms of models to test, I mostly use Crystal Clear XL because it's very good with old school art styles which I use for my DMing stuff.
@Cheru @ilovefist08437 Finally published the first version of the SDXL model I have been working on.
https://civitai.com/models/1072230
@severianvoid sorry for late comment. that's great to hear! Sadly, due to some irl stuff I can't check it out right now, barely generate images past months. But I will get to it when i have more free time!
In this comment I will just leave links to my character collections. In comments below I will provide some info on various portrait sets and why did I make them. Also why did I use such an old SD 1.5 model? Just randomly picked a model I had at the time xD
Some models will include real people, so be sure to set your settings to filter off X and XXX content, otherwise you will not see them.
CHARACTERS 1 - https://civitai.com/posts/16697937
CHARACTERS 2 - https://civitai.com/posts/16698426
REDACTED - https://civitai.com/posts/16734855
SYLPH - https://civitai.com/posts/16735041
LEAGUE 1 - https://civitai.com/posts/17096439
LEAGUE 2 - https://civitai.com/posts/17096507
CHARACTERS sets contain canon game characters. Some I didn't made because my SD model simply refused to make the stuff I wanted or because some mods already did a good job at it - for example, I didn't touch Sindrian Diktact because of the PAGSM mod.
REDACTED - yes, those are the AI cores. Red and glitchy Gamma ones. Stable yellow Beta cores. And beautiful and blue Alpha cores. I tired to put as many robot waifus as could - add Cortana, Demi, EDI, etc. Sadly, only so few have worked out.
SYLPH - I always thought that Sylphon RnD Mod lacks portraits, with those beautiful ships and, probably, my fav Large energy mount artillery battery of theirs. I tried to add some mad scientist flare to them and also match overall faction color scheme - white and purple with strange black stuff leaking from their ships. I wanted to add some feeling of illegal genetic experiments and UAC-esque delving into other dimensions.
LEAGUE - is for Persean League. Tried to give them militaristic vibes, with their navy blue and gold trim color scheme. Probably, one of the sets that turned out better than others.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.