Whats new in v2 beta?

MysticFrostwind Think it might be those "minor tweeks" to make it "warmer in color"..? The sample images do look pretty good, I look forward to trying it out.

yuichi69 not sure if you're into the furry stuff, but if you are, one thing to keep an eye out for, is where the D enters the A during gay sex in which the characters face one another. Cuz the main models are so heavily biased towards female, that it's very difficult to fully remove that when you don't want females in your generations.

That said, I've noticed recently, that with gay furry sex(and honestly, this may even apply to non furry, it's just more noticeable with furry, cuz ya see the tail coming up from the back, but then..), the D enters directly under the top guy's balls in the front (basically, into his non existent vag..). Whereas, it should be entering behind the top character, and be shaded from the front with only the base slightly lit (right, forgot to mention, I'm talking about POV images).

Honestly, it's one of those things that's pretty easy to miss, but I'm a sucker for detail, especially when it comes to male anatomy, since I was just starting to get decent at digital Art before AI went big. Didn't see much of a point of carrying on at that point. Too much work to get worse results.

But, I don't regret getting as far as I did, cuz it taught me a lot about the male form, shading, lighting, proportions, gradients, even a little about depth, and some of that has helped me to better understand how AI image generation works.

Oh, as per the point of telling you that, some models are better than others when it comes to more accurate male anatomy, even among the gay models. Just gotta experiment and see which one puts out generations that look the most accurate.

@yuichi69 Sorry for the sudden question. I've also been researching various aspects of the male skeleton, male body shape, shading, lighting, proportions, gradients, and sensitivity, and linking them to my own models, but I don't recommend simply merging checkpoints distributed by civitai or huggingface.

When it comes to male anatomy, simply merging distributed models will cause reactions that affect the learning process, resulting in poor results.

My best recommendation is to train Lora for the checkpoint and embed it in supermager.

The SakuraSakumaMix I provide only uses one or two models, and I simply input illustrations I drew myself for training into the checkpoint.

This should allow you to do this without excessive color overheating.

I hope this is helpful.

@zaizen "train Lora for the checkpoint"

So, you mean to merge the trained lora into the model? I've considered doing that. I use comfyui myself, and I know there's tools for it. It'd be sorta like fine-tuning the model, but with lower resource overhead.

@Lazman Yes, since anime models only output female figures, the method of teaching the model male body types and structures is currently considered the most optimal approach.

@Lazman dreambooth is also a well-known technique, but it's not practical.

@zaizen "anime models". Not just the anime ones.. The base models in general. The original SD/SDXL, and prob even flux, that all the models/loras are trained from, have a Strong female bias. Which is sexist as hell imho. If you make a unisex product that millions of people are gonna use, you don't just aim it at one gender..

Depending on the model that ya download, it can absolutely twist and warp any intended male images into being female. There's a handful of models that I've straight-up deleted out of frustration.

Dreambooth, that's the thing that comes as a tab in Kohya, isn't it? Yea.. I never did end up using Kohya. I tried to get it working but no matter what I did, it would OOM on me every time. Onetrainer was much simpler to get working, and is a standalone program to boot (no webui crap).

Unless I was just using it wrong (possible), I think Kohya may be for those that have at least 32+gb Vram(AKA: rich people). But, I've only got 16.

@Lazman Yeah, I agree. The mainstream models with high download counts have unbelievably cheap modeling, painting, and coloring.

I myself draw pictures from scratch without using AI.

That's why I was frustrated with AI and opposed it at first.

But since I'm good at drawing, I decided to teach my own methods to AI, and that's how I created SakuraSakumaMix.

No matter how much you emphasize (male focus) on mainstream models, only the hair and chest change while the overall sculpt warps into a woman—it was truly infuriating and unacceptable.

So I used my own drawings to teach it gender differentiation, proper painting techniques, and accurate color density.

I only have 16GB too, so I can't run Dreambooth.

I just trained one IL model using SuperMager like I showed you earlier.

It should be possible with automatic1111 and the Forge Git.

Sorry this got so long.

@zaizen "That's why I was frustrated with AI and opposed it at first."

So was I. I even initially vowed not to bother with it, But ended up deciding to use it out of curiosity, cuz the quality of some of the things it put out, was hard to ignore. And, I wasn't that great of an artist anyways. I was just starting to get decent, but always been too self-critical to be truly good.. I had difficulties drawing organically, and making things look natural.

At one point, I found drawing to be therapeutic, but my standards are too high for my skills, and it just became stressful, and almost depressing. Had I been in a better mental state, I may have continued with it.

But then, there's also the time investment vs value. I mean, when an AI image can be done in seconds, that would take me days/weeks, After years more training, well, I just don't see the point.

"No matter how much you emphasize (male focus) on mainstream models, only the hair and chest change while the overall sculpt warps into a woman"

Yea.. I know.. I've got my comfyui set up so I can view the image as it's generating, and even on my setup, having gotten it as good as I have and using custom loras, you can still see it repeatedly (during generation) adding and removing vagina.. and it always starts with tits.. and on the odd generation, it will still output a character with a vag.. /barf..

SDXL just thinks human = female..

And clearly women are the ones that need 'more' representation in modern society.. Oi.. And don't even get me started on the modern media..

Don't get me wrong, I got nothing against women in general, but female anatomy is absolutely disgusting when it's not something you want to see, and/or have gotten sick of seeing it everywhere.. And hey, if women can complain about dick pics, then how is a dude a misogynist just cuz he don't want to see them naked..?

Not that I've ever sent anyone dick pics. That's just distasteful imo.

"accurate color density"

I'm not so sure about all that. In my experience, Illustrious does pretty well overall, but, to be fair, my focus when I was an artist, was more towards shading, lighting, and anatomical correctness. So, I guess more composition than color schemes.

"gender differentiation"

Did you have any particular methods to teach it that? Oh, nvm, that's right, the lora merge thing. How bout keywords on the loras though, did you have any particular techniques, and did you train regularization images, or just the main dataset?

"automatic1111"

I haven't started up auto in forever. I mainly use comfyui these days. It's so modular, there's a lot that can be done with it. And I've never tried forge.

"Sorry this got so long."

Lol.. That's 'long'?

@Lazman Sorry for the delay in checking due to my heavy workload.

When I create base models, I take care to avoid making them look cheap. This is part of my professional expertise and role as someone who can draw without using AI.

“Gender Differences”

When creating LORA, I perform the essential task of “removing tags” to ensure the minimum necessary information is retained.

This is basic information, not a special method.

The AI sees the keywords in Lora through the images I draw myself.

1boy, 1girl. That's all.

“Accurate Color Density”

This is purely personal preference—what an individual finds beautiful.

Anyway, no matter how much you merge Wai or Nova, their color density and shading look pretty cheap.

I hated that and it made me depressed, so I used LORA training to teach checkpoints for color density and shading.

The result is SakuraSakumaMix.

“AI can create images in seconds, but humans take days or weeks.”

This is the inevitable evolution of our current era. AI will undoubtedly take over painting jobs.

But what if we focus not on work, but on living freely and unrestrained?

The reason I use SD and draw without AI is to “preserve my own artistic style as information genes.”

The unique style I possess is something only I can draw.

It doesn't matter if it's messy or chaotic.

My style is available as a LORA under the waryusyastyle name.

I'd be grateful if you'd give it a try.

@zaizen "When I create base models, I take care to avoid making them look cheap. This is part of my professional expertise and role as someone who can draw without using AI."

That's respectable. many people put out sub-par work/s, and it just kinda pollutes/dilutes the already flooded content pool.

As a young child, everything I gave to my mother, she would say is 'Great'(which, even as a child, I knew, couldn't be right), so with no other measure of reference, I became my own worst critic. And I think that also made me more critical of other people's works.

That, and I've always been logical(what with being autistic(though I didn't learn about that until age 38 or so)). So, I just don't see the logic in releasing sub-par work. I mean, it's one thing to Do subpar work if you're getting underpaid by a corporation.

But, when it comes to creative works; this is something that is representative of you as a person, and so, to release imperfect content to the public, just shows that the person has no respect for the craft, or that they're unintelligent, and actually believe whatever praise they get for outputting trash content.

Sorry if that seemed a bit repetitious or convoluted. It's 2am and I'm half asleep.

No worries if it takes you some time to respond. I put out quite a lot of posts/messages/responses on the regular. Also, I learned over time, not to get overly invested in people's responses. Cuz, people got lives, and jobs and such, and I'm just some random individual online.

"When creating LORA, I perform the essential task of “removing tags” to ensure the minimum necessary information is retained."

Well, that makes it easy; just use blank captions, lol.. But, yea, I think I get what you mean. Like, perhaps you use an auto-caption program, then make sure to get rid of all the '1girl' and such it wrongly tags boys with.

"1girl."

You train with female data? But, isn't that counter intuitive? Or, was that just a general example?

"for color density and shading."

But, how does it handle realism? I do anime images on occasion, but my general preference with AI, is to attempt to create images that haven't been done before. Like, realistic images of anime boys. And with that, I usually try to stick to models that can handle both realism, and anime.

One of the main models I use is WaiXReal. PornMaster, and RealityMaster are also good. PlantMilk is also good. The homosimile models are hit and miss(random).. They can do real well sometimes, other times, they lean too heavily into cut cocks, which I do not like at all.. males are supposed to have foreskin, but 100+ years of ignorance towards male anatomy, means that barbaric practices like 'circumcision' (AKA; genital mutilation), still exist..

"But what if we focus not on work, but on living freely and unrestrained?"

What are you, some kinda commie? I'm joking. That's absolutely how we should be living. But thx to systemic brainwashing and indoctrination, largely the fault of America, too many people actually believe that this is the way we should be living..

"“preserve my own artistic style as information genes.”"

That's smart. Cuz your model will eventually get fed into others. These things are all accumulative. Well, one potential exception though. When AI video cards with 32+gb Vram actually become affordable(No thx to Nvidia, I'm sure..), more people will likely move on to flux, which can't be merged with SDXL. And so, SD/SDXL may eventually be left in the past..

Well, Unless SDXL/Illustrious finds a method to surpass Flux, due to the fact that millions more people can afford to do active research on it, compared to flux, which is so big, it's hard enough just squeezing it into an affordable card for img generation..

"It doesn't matter if it's messy or chaotic."

Honestly, it's cute, and the colors are nice, but the lack of eyebrows isn't really my thing. I can see the appeal for some people, but I find the eyebrows really help to define the face, and character moods and such.

@Lazman Sorry for the late reply.

Thanks for checking out my work—I know my style isn't everyone's cup of tea.

It's old-fashioned and not trendy.

But that's okay—I'm just happy you took a look :)

@Lazman Oh, right.

Since it's based on males and differs from females, we also need the “1girl.” entry to indicate that this part should be omitted. That's how I handle it when learning.

@zaizen It's not the 'style' just that one quirk about it. It looks great otherwise.

@zaizen I think I get what you're saying (and I've considered it myself). That you put it in, because otherwise there's no tag to put into the negatives in order to 'omit' that subject type. The thing is, that's not quite how the training works. If you add '1girl' to the tags, then it will think that '1girl' relates to the image of which that caption file is for.

You see, SDXL (or branch of SDXL) is More than aware of what '1girl' is, and so, regardless of the lora that you use, if you add '1girl' to negatives, it will omit that.

Although, there is one caveat to add to that. Since, like, 85% of the training data in SDXL, is of females, many of which are scantily clad, and even nude, SDXL basically thinks 'human' == 'female', which is why, no matter what you put in the negatives, and/or positives, it will still generate females at random(even with everything else tweaked perfectly, and using the best available models/loras).

I've also noticed that with some loras/prompt combinations, some models that generated males just fine before, will suddenly start generating females. So, if you notice that you're suddenly getting female outputs, try swapping the model to one that's decently different from the one you're using. For example, if you're using 'nova-wai-homo', swap to a model that is not nova, wai, or homo, as they may carry on that same niche error in outputs.

In any case, I always encourage experimentation, so, do whatever you think will work, and always save lora itterations, and workflows that produce optimal images, that way you always have a reference point to go back to, even if you end up in a rough patch of generating nothing but bad quality images, and/or an endless stream of undesired female outputs.