Introducing lion-ish Maddox.

Hey you. If you appreciate my work, please do leave any feedback - even negatives are important, for they help me improve things I haven't seen myself or wouldn't ever notice them. Just a comment, an impression. No need to buzz, nor 5-star me. I'm doing this for I enjoy my toons, plus I'm brand new at this training neural-networks world. In the beginning I was just the digi-drawer, then tried AI and luv it. Now I'm merging both. AI aids me to speed up my pictures. But since AI doesn't know my own style (and I want only that, not against others at all.) I decided to "teach" it then. It's been a long interesting journey - sometimes treading in the dark - still I'm learning. That's why your feedback - even if just a few words - can help, motivate, make me improve a lot. Cheers!

On Toon Version!

WOHOOOOOO!!!!! Remember those reference pics? THose kilometric prompts? No more walls (unless you want to!)

And on that, check this post linked below, almost no prompts, generated on site and it still make it do. Desto, rest is your imagination! Have fun!

https://civitai.com/posts/991150

Then, tag has changed according to version, for this one it's "maddoxtoon".

MADDOX IS FINALLY HERE! I'd say this this is more a Facial Lora, but it work for full body too. Actually, it doesn't interfere with bodies anylonger. So, maybe any clothes, anything.

Gosh! Earlier versions were so NSFW that it gave me such a hard time to create this one. Heavy hard work!

Now can re-tell his story. Maddox was at first a pet (not even a cat, he was a HEDGEHOG! LOL) from a toon of mine - yesh REAL RPG = role-playing game for reals! After some time, on a totally different run, I needed a quick toon. It could be anthro, then first name which came to me was his. MAddox - the Lion. We played that game for a while. Short-lived. I even made a childish sketch of him. I like it, but I felt like transfering those traits to a human lad. That's when our Maddox of these days have been born. And that's the reason I was so obsessed with him looking how I idealized him.

This time I did everything right: Anime tags. Anime model for training. Hence, now, lol, Juggernaut Aftermath is not the best model for him, but any anime, semi-realistic, bara, 3d model will do.

Now he has real eyes AND EXPRESSIONS!!!!!

I am really proud of him. Although I've read that recent released guide by Euge_, had to use google translator as its written in Mandarin I think. They mention about the dimensions of pics on the dataset. I knew we could and even should user dimensions larger than just 512, what I didn't now is that the pic must be 512 whether width or lenght but both (when in a diff. resolution) numbers must be multiples of 8. And not all AR work. Hence maddox had proportion, distortion issues. Maybe that's the reason my OiMate XL, which I retired also had same problem.

This time it was only 3 diff. dimensions since dataset wasn't that small, yet not huge as it was hard AF gathering good pics of him with him looking the way he should, plus variations and the likes.

Clipskip 1. Anything. I tested with sde++ karras and 2m... but any of those conventional anime popular sampler should do. This time even when set at 1.0 pics aren't rendered out BURNT! Which means, LoRA is balanced now. For some of those faces I noticed some people luv, like a ginger-ish DILF, remember, he's the king of 1, 12, 123, 1234... seeds. Use them.

But set weight to 0.8 if you want more diversity. Adetailer if face doesn't sound. Noticed that theres no way, you will have to mention "cheekbones" every single time.

USe those usual quality tokens, embeds of your desire, must be compatible with other LoRAs, and even his previous incarnations if you want the "bod".

This one IS NOT NSFW! I haven't included any nakey for the prior versions had them ad nauseum! lol

Even without a single nakey, erec. under clothes and stuff, he still gets too horny. If you read the tag on Webui and see "ass" it's because I put a pic with him standing giving his back at us, but turning his head at side to look at us. Noticed that at least 1 pic like that improves a lot the toons variety (and "fluidity" for AnimateDIFF) - Learnt that with my J-Son, he excells at AnimateDiff if you know how to prompt, and his NSFWs there are... blushes and grabs a fan

Have fun!

You can generate him at Tensor.Art:

https://tensor.art/models/665203895125108723

ps. I do NOT recommend merging these different versions. Unless you really know how to set merging steps (even I don't lol), it will ruin it. You can fully skip v.1, 1.5 is practically the same but better (but overtrained, which means you'll need STRONG PROMPTs not to get NSFW 100% plus some "side-effects"... results from overtrainning.) Yet, rather not merge toon and 1.5 for it's quite likely "toon" will disappear and you'll get a broken 1.5... (not to mention activation tags are different). I guess you can merge v.1 with v.1.5 but it's likely to make it even worse I guesh... not sure. lol

On V 1.5:

Nope, not v.2 yet. And file name is maddoxmale-15.safetensors. Tag is "maddoxmale".

At first it would be the usual 2.0, then I looked at results... and thought "Maybe should call it 1.5 instead?" Again small dataset, but pics were good. I've noticed that it's more important lots of variety not only on poses but on backgrounds and clothes too. I have with v.1 gathered about 100 pics of him. But I wasn't satisfied with results, as I've said, on v.1 his eyes are overshadowed and we can barely see them, plus there are some stuff related to skin and their marks which the training absorbed which I dislike. Not to mention his facial skin, yeah, he has those "prominent" cheekbones, but their usual shadow had been taken as brown-skin, not even beard! Hence not matching the neither the whole head nor the body skin tone. But we have to start somewhere right?

Welp, what has changed now? ??????? is mostly the same, but YESH! NOW HE HAS EYES! I put blue eyes on the sample pics so that they are visible, but it should work with any colour (on his data he has even white eyes lol)

Why then am I saying 1.5 instead of 2.0? I usually tag the dataset with anime tags, this time tried the CLIP tagger, I really don't like it, for it lack details (maybe it's good for styles, but for toons...) then retagged with anime. BUT I THINK I MADE A MISTAKE afterwards. On all of my LoRAs I pick that anime model for training. This time, I thought of switching it and the chosen one was SD 1.5... for I thought it would look more realistic. I think it was a bad move, as it had absorbed exactly the things I wanted it to get rid off - plus it takes the dataset as anime and may render anime-eyes or weird ones (wouldn't say ugly, as they really look cat-like heh.) Maybe SD 1.5 got exceptional well the concept of him being a cat-like human-male and took it as the single and only rule for reals?

Desto, if you need some reference pics of him, have a look at this post here:

https://civitai.com/posts/884085

That's how he was supposed to really look like. From initial pics I posted of him, the best one IMO is that one where he's wearing camouflage jacket, gonna pin it to be the first pic. Even if not the most popular. But I don't like to tell lies, not lie about the things I do. I already do only what I want, hence no need to hide things.

Then, sorry for this, he does work, do not worry. I haven't even used ADetailer!! But be ready for negatives if you see things you don't like (malformed, deformed and stuff). If you're setting random seeds, if you like the pic in general, but it has fugly stuff, re-use the same seed but add negs. I'll have to re-train him... later.

Good side though is that now HE DOES HAVE THE FACIAL TRAITS I WANTED HIM TO! He is now the Maddox I imagined from the very beginning, first sketches and stuff. On that I'm truly satisfied, maybe even elated hehe. YESH! HE'S A FELINE-LIKE TYPE! lol

I haven't tested him that extensively, but from what I've seen, he works well with almost any model now - even realistic ones. But on realistics be ready to input negs about fem. clothes and the likes.

Still, he works well at 0.8, but I believe best results were with weight at 0.7. Clipskip at 1 (my LoRAs usually don't need to skip the 1st frame, even when models require that,) just in case pick any number of your likings or related to your model. Interesting though is that I've noticed that whether on F32 or very prunned and with clip-set not broken (or fixed) F16 (which means F16 but heavily hand-fixed) is that on my LoRAs you get best or interesting results setting clipskip at 1, 3, 4 and 8 (!!) and EVEN TWELVE!! Telling you the truth! So the usual clipskip 2 may not be the best one here,

On highres-fix - YAY!!! - now they don't, or shouldn't, remove his traits. Then maybe R-ESRGAN of any type... BUT FOUND IT OUT RECENTLY THAT FOOLHARDY may alliviate the rigid over-fit over-veiny over-muscled ???????. Also after paying attention at some pics, I've noticed he has on the images trained DOUBLED UPPER BODY! If you add to negative "unrealistic doubled neck shoulders chest ribcage..." your model will get it, and will try to get rid of it. And adding to positive "thick ossature male body" HELPS A LOT!

If you want to pair with other LoRAs, then maybe lower his weight, and put him inside ADetailer. For full-body pics, you might want do that too.

Hum... maybe his v1 for "BOD" and v2 on ADetailer for face? lol

On V1:

Note: Second batch of up-ed pics are part of his dataset for his v2. Some have been manually edited prior to post. Prompts I used were more or less like those, with him inside ADetailer - to force him have eyes, as original dataset is overshadowed, his eyes are deep, then they hardly show up without Adetailer. Don't take these pics that seriously. IT's mostly a showcase about what's coming next. He IS (was...) supposed to be brunnette black-haired but as you can see even I am unable to keep it that way. Models insists on rendering ginger-ish. Not that bad though as suits him well too. Shame his long beard is something for brute-force prompters (theres a LoRA or embed for t hat, which I don't have)

And yesh, besides head-shots, I use a lot of img2img to re-use the pile of pics here gathering dust. Speeds up the process. If you read EXIF you will laugh, as there are mixed "????" into it. Because for img2img i didn't care. Just anything.

Since he's part of Oi MAte! For NSFW dataset, I "equipped" such LoRA and set it around 0.7. lol Put it at 0.7, then Maddoxmale at 08, type a simple nsfw ???? prompt, DPM2, 22, cfg 7, seed 1234: there he is, quite tempting I bet! Ah, model SHOULD BE FIRST SD 1.5 RELEASE: that 1.5 ema only. (the smaller in GB of all of them, and must be the .chpt not .safetensor)

BTW, SO FAR HE IS THE KING OF SEED 123, 1234, 12345. USE THEM!

Another character of mine which I wouldn't ever share if he wasn't shared yet. He's part of Oi Mate!

But this time a Lora only for him. Reason I like MAddox is the same as another toon I have (and posted only on DeviantArt,) which is Rage. They are faces you have seen it already, but they appear rarely, mostly on core SD and then with all the things the core model brings: distortions, overshadows... On Rage I decided to simply grab the only ultra jagged weird pic I had of his face and repaint the rest and restored what we had.

Maddox is somewhat similar. The difference here is that he started from scratch as a sketch of mine. For as I'm pretty sure Ive seen him, but couldn't ever reproduce that same blurry image I had on my mind.

We could say that both came or are heavily inspired on whatever they have fed SD 1.5. Now all other toons of mine came from myself. Not saying they might not resemble any models in-built character, but for sure some traits they bring are intrinsic belonging only to them.

In our Maddox case, he's the typical "brute", rambo-ing in and out of the jungle. Strong cheekbones, yet a more "oblong" (had to do my homework to learn how such head shape is called) head type, not the default "square" which is - weirdly - the typical SDXL shape. Plus, as almost all my toons, a thick bulbous manly crooked nose which leaves no place for the , again, default, ultra-high philtrum. It's really hard to make models understand that.

Philtrum (the area or the "groove" between upper lip and nose\nostrils) is even bigger on generic XL, not sure why. I'm not against that shape, what I don't like it because, it is BIAS, it's a female head shape, on them, the girls, their (bias again) meaty lips need a higher philtrum so that those upper lips dont get into the nose/nostrils, but on men, it's not impossible, but it's less common - regardless human "racial" type. Philtrum also gets longer with age too.

Also many of my males have those "prominent" chin, which means their teeth would fit better those "retracted" teeth-line, barely impossible to render - as all the rest. I've seen models then rendering doubled-noses, chins... etc. I'm not sure if I added a real plain and simple whether those grey 3d shapes or facial bones (as a skelly) would help or would just worsen things. On my Oi MAte XL, I tried teaching those things, results came all wrong, instead of "learning" such traits, XL took them as "malformed". (I really need to learn how to deal with XL better)

Another part I'll have to figure out how to deal with is the hair. Have you noticed that, even Asian males come out mostly with "Pompador" Hairstyte when you ask to render short hair? That style is past now. Plus, my toons usually are planned (when as a concept) with very very short hair, like military, almost-buzzcut, short short short but messy, wavy... That part is hard to train too.

Maybe I should train a whole checkpoint then lol

Desto, not only this one, but prolly all of my LoRAs perform better on models able to 3d, for while rendering, the model is able to adjust formats. Realistics are usually only trained with "plain" photos (raw or not), hence their flexibility lessens (I might be quite wrong here, for at least until its mid-releases Consistent-Factor was very flexible, shame too censored on men, but nothing an extra LoRa couldn't "sort it out").

Now you know why even by swapping his face (with his original textures) results are subpar. I need to train his head, those facial bones structure and particularities. Hence, I'm training him. But to save me time. At first - differently from Atlas - I'm making him a "facial" LoRa only. So that I can extract more pictures of him and build a proper dataset, then a proper lora.

ALTHOUGH THIS ONE IS ALREADY Full-fledge in order to get him. Has mostly headshots and some nakeys. Among all lads, he's got the best body (not sure this LoRA) can be on par with data, pair with Oi MAte! for general pic, and put in inside ADetailer, just in case.