/siberia/ - Off-topic

Here's some goon material for those who want it, although I have a feeling that this one might be a tad controversial.

>>714858
well i enjoyed it anyway

Some more goon material.

In this house we appreciate hags.

secks

I've recently been testing an anime-style finetune of Chroma, which itself is derived from the more up-to-date Flux architecture. The results so far have been pretty promising because it has the uncensored dataset of Illustrious/NoobAI/Pony models alongside the ability to use both natural language and the booru-style tag system I'm more familiar with from the SDXL architecture in terms of constructing the prompts. While there aren't a lot of native LORAs for Chroma yet, it does seem to be able to make use of existing Flux LORAs quite well. Because it's based on newer architecture, it's also able to do mechanical objects like firearms a lot better than stuff built on top of SDXL's foundation. Pic related is a quick test I did with a straight copy-pasted older prompt I'd previously used with an Illustrious finetune model, a gun LORA and a Disco Elysium artstyle LORA. Minimal manual edits were required, mostly to do with the facial hair.

>>718632
requesting Karl Marx, Gilles Deleuze, Guy Debord and Mark Fisher in Disco Elysium style

>>718632
Could it create Starsector or other game portraits?

>>718634
I made this like 2 or so years ago so I'm sure something better can be made now.

>>718733
Looking through CivitAI, the latest I could find was for NoobAI, so yeah, I could probably cook something up in that style but it probably wouldn't look as good as something made with Chroma. The best solution would be to make custom LORAs for each subject (Starsector artstyle, Marx, and what have you) in the newer architecture and then generate with those for the desired results.

i mistrust AI but i still enjoy the silly shit you can do with it

Here's some art of my favorite Italian political figure.

>>718634
I tried to make one of Marx based on that one black & white photo of him.

More Jucika Art.

Ernesto Guevara if he was still alive today

>>719404
Looks cool, but I feel like he would not look that young if he somehow managed to live all the way to present day given he was born in… *Chekas notes* 1928. At the very least he should have some grey hairs and almost certainly more wrinkles.

>>719412

Sora 2 isn't available in my country so i get what i can get, run it again see if you get something cooler

i asked it to give another hand to this neoclassical Kali it fucking got wrong the first time

AND IT GOT IT WRONG THE 2ND TIME LMAO

>>719403
Please sit on my face!

Something other than Jucika for a change. I'm really taking a liking to the DE artstyle LoRA and how versatile it is in terms of application.

Some guy on Twitter asked me to make a picture of Jucika posing in front of János Kádár, and I felt inspired enough to make it.

>>719292
>>714942
god why are men with firm bellies so inticing to women?

>>719635
uoh secks
>>719675
also segs

>>719635
>>719675
Do Alunya next

>>719777
Checked. Trips say I gotta do it, so I guess I'll make my best attempt at it.

Okay so here's Alunya. It came out surprisingly well I think, despite the fact that I only used a style LoRA.

>>711119
> it tries to mash multiple different gun models together when all I want is a glock.
can't you train a separate model on guns and merge it with your pic gen model

(I have no idea how anything works)

Here's a more SFW Alunya piece for Caturday.

>>719939

AI sucks for making guns

3D gun nerds have to rely on other gun nerds constantly pointing out stuff we missed because finding reference for this stuff is usually such a pain.

Had to delete the old version because I realized that it had a bunch of errors that I didn't notice the first time around.

>>719939
It only works up to a degree in that it will get the overall silhouette right, but it will still screw up individual parts (such as the fire selector on an AKM), and all SDXL-based LoRAs that are available publicly are specialized for making one specific kind of gun. Flux Loras are better at following prompts and generally just more accurate, so if you have a LoRA that makes most common types of guns in general look better and specify that you want a 1911, you'll get about the result of what I made with >>718632, where the mistakes are so minor that you can easily edit them out by hand.

>>720076
This is also still true though, if you want your gun to look right, you still gotta know guns yourself to be able to both spot and edit out mistakes. It's why I recently visited a museum that had a display section full of old weapons all the way from muskets to the first half of the 20th century.

>>720086

i'm a huge gun nerd, but out of memory you can only remember so much.
which is what blueprints and references are for.
one misshape and you'll have every commenter and their mom saying how the angle is 30° off

>>720078
what are your text prompts like to get the specific results you want

>>711117
>always provide sauce to credit the original author.
Absolute state of artist faggot protecting their private property.

>>720171
Depends on the model, what infrastructure it's working off of, and what you're looking for. With solo pictures like the Alunya one, it's easier because you only gotta worry about the details of one character. With multiple characters, you gotta stop and think about formatting to avoid having details bleed in between characters (something I struggled with early on). The more specific you are with your prompts and the higher the CFG (Classifier-Free Guidance, basically how closely the model should follow the prompt), the less randomness the model will introduce when it's in the sampling phase of its generation process. However, especially with SDXL-based models, there is the issue that if you crank the CFG too high, combined with the wrong kind of sampler, it'll just result in your image having eye-bleedingly bright contrast, which you then have to fix by adjusting the color values, brightness, and so on in Krita's native filters.

Just as an example, here's the prompt I used for the Alunya piece.

Style Prompt (This is a Krita thing, it's mainly for if you have a style preset that you've set up for a specific workflow): expert shading, masterpiece, best quality, amazing quality, absurdres, very aesthetic, 5toes, 5fingers, {prompt} (←- This is where the rest of the prompt starts from)

Common Text Prompt (basically the one that applies to the whole picture, this is where I usually put stuff that's in the background):
sfw, abstract, disco elysium, expressionist, harsh brushstrokes, painterly, traditional media, outdoors, urban, cityscape, park bench, trash can, dirt road, grass, tree, motion lines, sunlight, white wooden sign, metal fence, english text written in big bold red letters reading "NO LEANING ON THE SIGN!" and "(FINE UP TO 100$)" written underneath in smaller text,

Regional Prompt (Basically what you apply to, like an individual layer or layer group): BREAK (←- This is a way to separate sections of the full prompt manually to avoid bleeding) 1girl, solo, full body, catgirl, toned female, biceps, black cat ears, pink inner ear, black cat tail, short hair, black hair, bob cut, red eyes, eyelashes, thighs, hips, breasts, choppy bangs, thick eyebrows, pointy hair, matte skin, teeth, fang, smug, open mouth, smirk, standing, facing viewer, leaning on white wooden sign, leaning forward, standing on one leg, looking at viewer, black dress, pleated skirt, short sleeves, red belt, black boots, pointy red and black bandanna around neck,

Negative: Empty (Models based on Flux-architecture don't use negative prompts anymore)

>>720227
Thanks! That's very interesitng. I didn't know Krita had AI stuff embedded in it.

>>721259
>I didn't know Krita had AI stuff embedded in it.
It's less that it's something officially embedded into the latest builds of the software and more something you CAN use if you just install the Stable Diffusion Add-On (found here: https://docs.interstice.cloud). I would recommend running a separate ComfyUI instance through stuff like Stability Matrix over doing the auto-install just because the CivitAI integration makes it easier to download and install new models, but if you want just the basic stuff, the auto-install of all dependencies via the "Local Managed Server" works fine.

Experimented with making a piece with more than two people while trying to keep their features from bleeding into each other. It took me a few attempts and a whole lot of trial and error, but the end result speaks for itself I think. Turns out that the secret was in dividing the picture into sections with masks and then assigning each section with its own regional prompt that in turn was associated with a character.

Gomrad Spurdo, where did I go wrong with this prompt:

<Karl Marx, a tan German male with a black moustache, gray hair, gray beard, wearing a 19th century suit, shaking hands with Adam Smith, a Scottish male with a white powdered whig and wearing an 18th century suit, while Deng Xiaoping, an elderly Chinese male with short black hair and a 20th century Zhongshan suit, smiles in the background

>>723190
Depends on a lot of things, firstly:
>What base model did you use to make this?
>Did you use any LoRAs or ControlNets?
>Did you run the process locally or through an online service? If it's the latter how much control did it give you over the generation values?

Then secondly, the prompt itself is currently written as one continuous string, which you should never do if you intend to have multiple characters featured in a picture at the same time. You need to separate them, or otherwise you're going to have features of one character bleed into another.

Thirdly, there's the issue that you're trying to generate specific historical characters, and the model likely doesn't know exactly what they look like by default because base ImageGen models are mostly trained on wide-ranging datasets and common concepts, hence why you need to use ControlNets and LoRAs for more specific things.

Also, on a side note this Ghibli-ass art style makes it obviously look like something made with ChatGPT, which is just… ew, no.

>>723211
>What base model did you use to make this?
https://huggingface.co/spaces/prithivMLmods/FLUX-LoRA-DLC
I used the one called Ghibli Flux

>Then secondly, the prompt itself is currently written as one continuous string, which you should never do if you intend to have multiple characters featured in a picture at the same time. You need to separate them, or otherwise you're going to have features of one character bleed into another.

ok thanks

>Thirdly, there's the issue that you're trying to generate specific historical characters, and the model likely doesn't know exactly what they look like by default because base ImageGen models are mostly trained on wide-ranging datasets and common concepts, hence why you need to use ControlNets and LoRAs for more specific things.

gotcha

>Also, on a side note this Ghibli-ass art style makes it obviously look like something made with ChatGPT, which is just… ew, no.

i thought it might fuck up less if the art style was simple

>>723211
> You need to separate them, or otherwise you're going to have features of one character bleed into another.

how to separate the description of two characters while still acknowledging they are supposed to be shaking hands? is this complicated also by there being a third character?

>>723211
lol it made them all asian. here's the new prompt separated into multiple strings:

>Karl Marx, a tan German male with a black moustache, gray hair, gray beard, wearing a 19th century suit.

>Adam Smith, a clean-shaven Scottish male with a white powdered whig and wearing an 18th century double-breasted coat.

>Karl Marx and Adam Smith shake hands.

>Deng Xiaoping, an elderly Chinese male with short black hair and a 20th century "Mao" suit, smiles in the background.

I'll look into ControlNets when I have more time.

Posadist meme

>>723258
>>723261
If your intention is to make anime-style stuff, but not use finetuned models running on older SDXL architecture like Illustrious or NoobAI, which are easier to run locally (That stuff only needs like 6GB of VRAM. You can get that sort of performance with even a decade-old GPU), I'd recommend this one instead:
https://civitai.com/models/2022057
It runs on the same model architecture as FLUX, so it can use FLUX-based LoRAs, but it's finetuned for anime content, unrestricted in terms of what sort of stuff it can generate, has lower hardware requirements, and can be prompted with both natural language AND booru tags. It's what I used for making both of the Alunya pics.


>>723259
Regional prompting (if available) is the best solution, but if that's not possible, alternatives do exist. Using separators is one such alternative (this is what I used initially) if you can only use one positive and one negative prompt, in which case things should be formatted like this:

Initial style and medium,
BREAK Character 1
BREAK Character 2
BREAK Character 3

I try to avoid using models that rely on natural language because that can be kinda finicky. I tend to use tag-based ones instead because it's what I'm already used to working with. Also, for the handshake, specifically add into both of the characters' prompts a part where it makes it clear they are shaking the other's hand. Something like "Marx is standing on the left side of the picture while shaking Smith's hand" and vice versa for Smith, the important part is that the model knows how the picture is supposed to be, because otherwise you're kinda just rolling the dice on where the characters are gonna be placed.

The most important rule of thumb when it comes to prompting is that you should always prefer to be more specific over being vague, because if you're vague, then that leaves things open to interpretation for the model, and it will inevitably interpret wrong because, at its most basic, it's a probability algorithm, not something that can understand author intent. Never trust the model to know what someone looks like based on name recognition alone!

>>723267
Nice meem.

>>723277
To add to what I said about the SDXL architecture-based models: They're gonna be the ones that (currently) allow you more fine control over the result if you know what you're doing. FLUX-based models do not currently have anywhere close to the amount of community-made LoRAs or fine-tunes of the base model as the older ones do. This is mostly because, surprise surprise, the older SDXL-based models with lower hardware requirements and the larger existing base of users are gonna have a bigger pool of community-made tools and material for you to work with. The FLUX-based models like Chroma do have the advantage, though, that they're better at following the prompt you put in, don't require you to use a long-ass negative prompt, and are more forgiving in the sense that they don't produce horribly oversaturated images if you fuck around with the generation values like CFG or the number of steps or use the wrong sampler, scheduler, or VAE.

>>723277
>>723293
thanks gomrade

>>723293
did you make that picture? its cool

>>723319
Nah, that's just an actual screenshot from Black Lagoon. Pic related, on the other hand, is mine. I made it for a fanfic of mine.

New Jucika Art. I tried to go for a sort of "folk dress" look with this one.