<h3 id="heading-30184">b33r b3lly now - or just get normal bodies!</h3><h3 id="heading-13">No, this is not a joke, is it?</h3>Eh, I mentioned that someone asked me to make a LoRA on "beer belly".¹ I had other things to do, but felt like making it - so that I could "apply" and test what I've learnt so far, and it seems it's not bad, not at all. Details far below.<h3 id="heading-30185">What does it do?</h3><ul><li>At higher weights it can lazily create pics of men with big bellies - aka beer bellies. Laziness in exchange of freedom. At least poses aren't bad... heh</li><li>Lowering its weight can still create beer belly-ed men (haven't tested on women or whatever,) but it requires some prompting - increases versatily and stuff.</li><li>Lowering its weight or depending on the model can also just make men having normal bodies, without being the usual "over-fit" - although while it does it, it wasn't the focus of this LoCon, yet since I've seen some people complaining on that...</li></ul><ul><li>Info update:<ul><li>usually SFW (depends on model)</li><li>tested and works for any gender and humanlike anthros (see sample pics)</li><li>at +0.8 poses get stiff as carboncopies of dataset, hence</li><li>keep at that or lower it a bit</li><li>noticed effects only at 0.3 upwards</li><li>usually at 0.4 ~ 0.5 bellies are just unfit without 6pack</li></ul></li></ul><h2 id="heading-42">Is it SFW or NSFW?</h2><ul><li>Both, mostly SFW (I think)</li></ul><h3 id="heading-43"></h3><h3 id="heading-30186">Why should I use it?</h3><ul><li>There are other tools available on Civitai. And sometimes just prompting is enough. Hence, up to you, maybe you shouldn't? heh But...</li><li>In case you want it just because yes</li><li>It doest matter if it's the best or not, as you cannot set sliders weight on your ui, hence rather this then nothing</li><li>In case you feel lazy and just want to hit "generate" and whatever... (I have not tested it that extensively yet...)</li><li>You can also try this Beer Belly Lora too or instead. I found it out when posting this one and there were one model with a tag "beer_belly", and when I hit the tag, it took me to: <a rel="ugc" href="https://civarchive.com/models/15372/beer-belly-beerbelly-v1-lora">Beer Belly (beer_belly) v1 LORA</a>. Belly buttons seems to look better there and it doesn't make your character muscular if you prefer that. :)</li><li>In short: you shouldn't. Then I cannot be blamed for anything... hehe</li></ul><h3 id="heading-30227">Okay, skip unimportant crap, go straight to settings, tips...</h3><h3 id="heading-30229">Lazy settings for lazy (and/or lazy purposes):</h3><ul><li>Weight 0.8 ~1.0</li><li>Clip: 1, 2, 3... 12?</li><li>Sampler - According to your model of choice. If your rich and powerful, try those new 3m which require huge amounts of steps (I don't like, me likes fast). Or those weirdo LCM, whether LCM sampler per se CFG 1~2.5 Steps 4.5~8, or LCM LORA (usually at 0.5~0.8) DPM++SDE (Karras) CFG 1.5~2.8 Steps 5~8, Euler A CFG 0~2.4 Steps 2~6, DPM2 CFG 0~2 Steps 1~5 (CFG at ZERO tends to ignore your negative prompt or if Lora check if reads the clip or not, if not, it ignores negative and part of positive too, especially if too long, they are meant for animation I guess)</li><li>HighRes Fix - Use the one which suits better your taste. Since we are at lazy mode on, steps 5, fix 15~25. If LCM not necessarily always UltraSharp, Just keep it at max 10, steps 15~20. If too dark, Clear/Clean Vae, but HR-F sampler should be (R)-ESRGAN or SwindIR steps 10 fix at 0.24~0.33 to deal with texture holes/pops on the pic caused by the vae.</li><li>VAE - I hate HR-F an VAEs too. Not sure on your end, but on mine, everytime models and stuff update pics look darker and darker. Adding more details worsen it even further. All VAEs are biased in a particular way, whether on colors, on genre, on etc. And many amplify the generic AI image look, but since this is laz-z-z-z-e-y mode... CLEAR/CLEAN VAE and R-ESRGAN_4x steps 10 with fix 25. If white texture pops still show up, increase only steps maximum by 5 each time. Since we're lazy we keep highres working on all pics. Fixing even those we will discard. lol</li><li>Positive prompt - b33rb3lly alone for whatever. DO NOT TYPE "BEER BELLY" otherwise you will, yesh, prolly will get the belly PLUS beer (pint, bottle, etc) somewhere in the pic (unless that's what you want) Other tags which work are dilf (lol!) man, bara. DILF and BARA are mostly a joke, bara is sorta universal, Despite being tagged as DILf, many 1.5 models have no clue what that thing means (yet they what milf is, unfair) Then DILF on its on may not work, but hey this is a LoCon, and lazymode, 0.8 and whatever! To make sure your male will be fat or beer-belly-ed whether belly and/or fat. SD tends to take clothes as skin/body, hence for fat 24/7, try to pair fat with clothes. Male clothes have a very few options, most descriptions are ignored where they should be targeted at and bleed onto the rest of prompt (opposite is true), hence "blue eyes", insta blue short, "pale skin", insta white clothes. "ginger hair", insta red t-shirt and maybe orange/reddish bracelet. "green eyes very open" insta surreal weirdo strong green lighting ray/beam/whole scne... "thin mouth" insta giraffe long neck and body, prolly thin.</li><li>Negative prompt - "taut top" to get rid of those weird tore fabrics over shoulders only... and real taut ones (pullled shirts may or may not vanish either.) Check the sample pics. I was lazy AF and it worked. heh - If you get a shirt "explosion" with a clear circle around the belly, try these "scar marks tattoo at the sides of belly". It's been trainned to deal with that (won't say it will work flawlessly every single time, but on testings I got one of that, then grabbed the same seed and added those tags in the negative prompt. It returned with an open jacket with correct perspective:1.3. ;) And on that, honestly, not sure it's this LoCon which does it, or the model, or both together... We should unite and make a mal</li></ul><h3 id="heading-30230">Recommended settings for creative and "creative" purposes:</h3><ul><li>Weigh it at 0.8 and give it a go. The lowerest I've tested it was at 0.3, curiously enough at 0.3 bellies are bigger than at 0.4. Tested same propmts without Locon, especially on simple/short prompts I've got a nonsense image with Locon off or below 0.3 - but still I'd go for 0.4~0.5 for normal bellies, 0.6 it gets bigger and bloated AF after 0.7. Hence 0.6 ~0.7 for very creative poses and scenes. Or 0.7 and open pose...</li><li>If you just want to try to get a "normal" body, then don't add "belly". As in some samples above, just by lowering the weight (.4~.5), and mentioning "b33rb3lly" (or nothing) plus the rest of your prompt, may work. I haven't even described anything on those! LOL</li><li>Most of small dataset is, unfortunately, whether indoors or under a ceiling. That's the reason I opted for a LoCon instead of a LoRA... then, just add "outdoors" and "yadda yadda background", on POSITIVE! If you put "ceiling" in the negative prompt, chances are it will simply crop the head out of the frame!!! Since I hate "poser" poses, most of them are very simple ones with "arms at sides" with hands down (or inside pockets), one or another is at head (i hate those, due to high-frequency of deformations), there are NO interlocked fingers, locked hands, locked arms in the dataset. And most of the time hands and fingers ARE NOT CLASPED and there are NONE holding any prop. He is standing on all of them - but I knew he would adapt according to prompt - it's not over-trained nor under- either! YAY!</li><li>Faces are somewhat generic... I focused on the belly - makes sense right? But it's likely any facial "addon" (LoRA, extensions...) will do. Lowering the weight increases freedom for different subjects, but may require extra input on the anatomy.</li></ul>If you start to get ABS (6- 8- 12- n-pack) on the belly, whether increase LoCons weight or work on the prompt - or both! Still I know this .safetensor is not perfect (flawless:1.3). Then it may fail. Sowy - but I swear if it happens, it's the checkpoint which is stubborn! lol<h3 id="heading-30231">The story...</h3>¹ I've mentioned on one of my LoRAs that I'd been asked to train a thing for these belly-types. Then I began with simple sketches fed into AI, but results were very subpar. I thought I was smart and would quickly and easily gather my dataset by "collecting" pics from the web - 'twas then when I faced the reality on the "male anatomy is harder". My arse. Not the anatomy, but finding good source is the real struggling part. Men simply don't know how to take pics for us to use them as dataset! Period! LOLI called it quits. TILL the new control net! Actually Ip Adapter!!! BOTH!! Then I began feeding it with my own material - which was very limited in quantity on such theme. First was booooh... Luckily on the second picture, it clicked! And I was able to extract those bellies I needed.I could have focused more and put more effort into it... but... I was also eager to test the trainning part and put into practice new things I had learnt. Then since, as YOU know, we rarely get perfy pics in the first go, I didn't want to move my hand to manually edit them, nor waste tons of tries on inpainting, I only used those which HighResFix could actually FiX. Once fixed enough (and by enough is not really enough, but...) I aimed at about 20. I got 18 (19 to be exact, but 1 of them doesn't count... because it was the only one with different dimensions.) But then, why not, I copied one which I liked best TWICE! lol Which increased my dataset to whopping amount of 21 pics.The End.<h2 id="heading-30232">You, enjoy!</h2>

Files with SHA256: f30736565ff82bdc616fa0a97aeea4948fda89496fd0d2c7b2fcefd210493cef