Just created this LoRA whose toon I already had who was just a side-kick on other LoRas of mine. Not to mention he served as base for many toons of mine.
Lantzer has now his own LoRA. Which is great!
I trained him on Colab (Pro) with settings provided by Victor Hall which you can find on Github.
Note: Sample images were uploaded because via workplace 'twas bugging out here and I could only add 3 samples. Images without base model were generated on Juggernaut (the main one is from older version on site,) or Redmond Liberte (on site too) or YES! S D 1.5!! (I still love it!) The others are in the post below.
It's a very slow simmer... He should be easy to use as the others. (I'm writing this while I'm getting his first epochs... Epoch 3 looks good - I only save each 3 or 5...)
Despite what's been written bellow, he has TONS OF PROMPTS to play with. Despite dataset being somehow small, quality used was honestly very good with different clothes and poses and facial expressions and some variations on type of hair. Not to mention backgrounds were unique in each picture.
Trained with excellent COGVLM captions, but I also added some Booru and some weirdness from BLIP. Which means that you won't really get a stable face all the time, but tons of versatility if you have creative prompts.
Since I followed a different method, he is almost underbaked, and depending on you weight him at 0.6 up to 1.0 - It goes well with any other Loras (which I experimented so far) without the typical overburnt effect - if it does, lower weights, cfgs and steps. People only overkill on settings on images they previously made as "sketches", you don't that on the first go. His textures in general (skin, clothes, hair etc) are very smooth, so no reason for blurred lowres textures, unless you're overdoing or it's your model. Just in case the "jpeg artifacts" negative may alleviate it. He shouldn't need negatives for hands as I hand-drawn most of them to look as good as possible. Instead of an embed, just add to neg what really bothered you and it should do it.
I've been playing with him on many different types of models: from real realistic to semi- and anime and whatever. Course results vary a lot, but even I am surprise on how flexible it is - it feels more like a LoCon than a LoRA, although as I've read LoRAs should really act that way. Anyways right or wrong, it is really fun!
I used 36 images of him only. And it already took me almost 1 hour of training. He has more than 300 images on his own. Although all those 36 were carefully hand-drawn and/or hand-edited on Gimp and on Krita. Images had the smaller side with at least something around 1500px and all of them at minimum 300dpi in truly crispy clear with NO ARTIFACT, most are truly smooth and they are a bit brighter to balance models which tend to generate absurd unnecessary dark images. What I've seen is that it's not enough, although it generates a nice contrast with vivid colors. And shadows aren't enough to make character disappear - although add UHD and softglow to aid somehow - or use any lora to control-lighting. In this case I'm not sure ClearVae is the answer as dataset is truly very colorful and ClearVae may fill the image with textures holes (or green fake textures). Not sure. Try and see.
His default is ginger facialhair/beard, hair is ginger too varying from buzz cut, bald and very short undercut. But any model can change it and such variations are actually welcome for versatility. In general I am very satisfied with results.
He is tall, and his body varies from almost slender to extremely bulky. Hands used were always free and arms usually at the sides - both visible - and never an arm behind/at head to avoid post-deformations. He has some facial portraits, some cowboy shots and lots of full-body images with some varied poses, although I didn't add intense action to avoid malformations. But it seems that he's able to adapt - and there's always controlnet for that (or just lower the weight a bit.)
On clothes and styles, he's more urbanwear, beachwear, sportswear, tracksuits (not bodysuits to avoid clothes rendering underskin then cropped), and even armors (masculine shaped armors only), cyberpunk style (with jacket and stuff). Here and there 1 or 2 with anything extra like a hooded jacket/shirt and a sidearm at leg. Lots of trendy masculine sneakers. All clothes were in different colors to avoid repetitive generations - yet it can happen if you don't declare which type of clothes and their colours you want (which many times is useless as SD tends to bleed out colour info to other things in compo.)
He is SFW - the extreme nudity is only fully topless and barefeet and legs - but always wearing shorts or trunks. I'm just mentioning it because this is the default. Any extra is beyond me! :)
Note: NONE of his shorts/trunk/lowerbodywear has laces nor cords for I'm fed with those. NO STRAPS. NO BAG NOR PURSE NOR BACKPACK. (You add them via prompt if you want them.) None of his shirts have nipples nor look like compression shirts. For I think the user should be given the option to pick the style they very wish. YET as I've seen myself it does happen, especially nipples. Ah, no clothes are over-creased for during my journey I've learnt that training models mistake those creases for objects and then ARTIFACTS! Yet, yes, while generating I've noticed creases but it's fine. And this is not a rule, but see if that's what you want when using any "add more details" as they will tore his clothes - not all LoRAS! - add nipples and even makeup. If you like that, keep it, but if you don't, just know where it MAY be the source. Models or high steps/cfg/highresfix do that too. Fact is that usually masculine LoRas are already very detailed... but to each their own.
This project has been shared only on this site. Just because yes.
He has been trained on clip 1. Hence try which works better for you. On Juggernaut he did better with clip at 1 (or zero on comfy.) On Anime ones, clip 1 bring more fidelity to his head/face look, clip 2 adds more variations. Hence up to you.
On HighResFix steps at 5 and fix at 0.3 should be quite good. Tried UltraSharp and AnimeSharp and they were good as well, but again, 5 and 0.2, I don't like going further than .2 because they tend to add make-up, fake eyelashes and lipstick, but again, up to you.
I usually generate them with
LOWER STEPS such as 10 and maximum 15
CFG 4 ~ 6 depending on model
Samplers DPM SDE++, SDE ++ Karras or fewer times with Euler (A) and LCM too. Other samplers only seem to add jpeg artifacts or simply burn the image - although I haven't tested which setting is the optimal for each of them. Unipc does it at 10 and lower cg too.
And I only upscale them in 1.5x to avoid extreme modifications regarding the original image.
Ultrasharp, Foolhardy or R-ESRGANx4 with 5 steps and HighResFix at maximum 0.2
More than that I manually upscale and sharpen the images on Krita which is always great to remove jpeg artifacts.
