Remert/Sniperstalker has an <a target="_blank" rel="ugc" href="https://x.com/BigRemerting">X account </a>as well as a <a target="_blank" rel="ugc" href="https://bsky.app/profile/remerting.bsky.social">Bluesky</a>. You can support the artist by <a target="_blank" rel="ugc" href="https://bsky.app/profile/remerting.bsky.social/post/3ldlowkazuk2c">asking them for a commission</a>, as far as I know.<h1 id="remert-legacy-09x78xp9l">Remert Legacy</h1>My main goal was to train a style LoRA that doesn't create extremely large, hyper-muscular, very narrow-waisted characters compared like the <a target="_blank" rel="ugc" href="https://civarchive.com/models/517833/remert-styles-2-vers-ponysdxl">more modern style LoRA</a> and only adheres to more realistic body proportions on default. I have also sorted through the dataset to make it so it's possible to create specific characters (with varying rate of success).<h2 id="general-usage-9sfz3j7w2">General usage</h2>This isn't a must-use guide, you should experiment with it on your own to find what works best for you, but my general tips/findings are:<ul><li>Adding "flat color, unfinished" and putting things like "detailed lighting" in the negative prompt can help in recreating the early look. It most definitely will not work when the word "detailed" is in the main prompt. I assume this has a chance to not work if there are other embeddings and LoRAs added. It also removes detail (duh) so keep that in mind.</li><li>Despite my attempts, on some occassions you will get a character with a narrow waist, except the effect won't be extreme, and adding "narrow waist" in the negative prompt will not do anything. It's just how the style is like, especially with muscular characters. If you include tags like "big muscles, big pecs, thick arms" in the prompt you can also boost that effect (bigger muscles = thinner waist). I found that using character tags (by including either "tf2", "left4dead" in the prompt or just generating specific chars) helps mitigate the effect the most.</li><li>It's supposed to be designed specifically for <a target="_blank" rel="ugc" href="https://civarchive.com/models/257749/pony-diffusion-v6-xl">Pony Diffusion V6</a>, but (at least from my knowledge) the general rule of thumb is that it should work with other SDXL checkpoints as well.</li><li>Recommended weight of the LoRA: 0.8-1.0 (but obviously if you use more than this style LoRA, you should probably lower it even more)</li><li>CFG Scale: 7 (6-8 should work nicely as well)</li><li>Any sampling method should work well, my favourites are Euler a Karras and DPM++ 2M SDE Karras</li><li>Steps: 25-40 (in my experience 27 or 37 works best)</li><li>Clip Skip: 1-2 (but in my experience I never found any difference)</li><li>During my testing it liked to draw hands with four digits or with very little detail on them, but putting "five digits" in the positive prompt and/or putting "four digits, missing digits, extra digits" in the negative can help. I recommend using adetailer or manually fixing the fingers on your own even if that doesn't work.</li><li>Faces in general usually look bad when making images with keywords like "full body, low angle". I don't use any upscalers so maybe that's my problem. Again, I highly recommend adetailer. I also recommend adding "detailed face, big eyes, detailed eyes" in the prompt.</li><li>Back muscles can sometimes look disfigured and unsymmetrical.</li><li>Should work best with human characters. Furries might work, but it still could draw human skin on the character.</li><li>It has a trigger word - "r-l", but when prompting for a muscular, masculine male character with body hair you it most often works without it. Try experimenting with and without the trigger word.</li><li>"Sunglasses" on their own create glasses with an orange tint. Just use "black sunglasses" if you want actual glasses.</li><li>If prompted correctly, it should be able to make SFW images as well. I find that most commonly I get generated images with nudity and out of frame censorship.</li></ul><h2 id="specific-characters:-06lzngdjo">Specific Characters:</h2>As an experiment, I've tried to tag the dataset in such a way that the LoRA should allow me to generate some specific characters. A lot of them simply don't work because they only have one or two images with them. I'm gonna include the ones that work best as trigger words. Most noteable (those that work) are:<h3 id="team-fortress-2-characters:-4ab6le6oy">Team Fortress 2 Characters:</h3>(You can use the "tf2" keyword to specify all of them at once, don't know how useful it might be, though.) Sniper, Soldier and Heavy might work best. Engie - not so much, sadly.<ol><li>Soldier - soldier_tf2 (I recommend adding "helmet" and "covered eyes" for the helmeted look, but it might sometimes give him a cloth over his head. Adding "construction helmet, engineer" inside the negative prompt also helps.)</li><li>Sniper - sniper_tf2 (you should add "sunglasses, cowboy hat", the australian hat he commonly wears is called a slouch hat, but my auto-tagger didn't recognize it as such so using it will probably make the character slouched)</li><li>Heavy - heavy_weapons_guy (generations with this guy often have a lot of contrast for some reason, should include "bald", also don't add "abs" if you want that chubby look)</li><li>Engineer - engineer (try including things like "bald, goggles, hard hat")</li></ol><h3 id="left4dead-characters:-xqqvbjwhb">Left4Dead Characters:</h3><ol><li>Francis - francis (will need things like "left4dead, buzz cut, short beard, goatee, tattoos" to look properly)</li><li>Hunter - hunter (but this probably only works well with "hoodie, hood up, pale skin" or "naked jacket" included, skin color varies between pale white and pale green)</li></ol><h3 id="naruto:-rocnih19p">Naruto:</h3><ul><li>Shikamaru Nara - shikamaru (will most probably need a proper character LoRA to look good, he was always drawn with small, dot-like eyes)</li></ul><h3 id="resident-evil-characters:-d0l08keve">Resident Evil Characters:</h3><ul><li>Lucas Baker - lucas (sort of works... I think...?)</li></ul><h3 id="detentionaire-characters:-2gqrasplx">Detentionaire Characters:</h3>In general, these (duncan, biffy, trevor, steve, emmett as well as texas from motorcity) don't work very well, if at all, but putting "detentionaire" in the prompt can give cartoony faces.I think for most of them a separate character LoRA would make the outcome a lot better, so if a character doesn't work for you, add another LoRA.

Remert Legacy - Sniperstalker/Old Remert Style [Pony/Illustrious]

Re-trained for Anima (super-rushed. half-assed edition). Tested on HomoSimile Anima and Uranus Vir Anima. (Uranus Vir gave me better results than on HomoSimile)Character tags don't work so well anymore. As in, they almost don't at all and/or get the characters confused. Not sure if it's overtrainer or undertrained.Example images have been generated on my own rig using ComfyUI.Use an upscaler for best results.Feedback is much appreciated.

Files with SHA256: 6217b8793e3524d325abb7c01d98314dfb76d1362555d1a306f2a9aaa67761f2