[Illustrious 0.1] Mesmera | The Biggest MultiStyle+MultiConcept Hypnosis LoRa Trained Yet

Mesmera is a LoRA (LyCORIS/LoCon) trained to enhance the capabilities of Illustrious 0.1 (and it's finetunes ofc), in generating hypnosis/mind control-related scenarios.Tested on these Illustious-based models: MiaoMiaoHarem v1.5a, NtrMix vXII, WAI_NSFW v9, Hassaku v1.3A.Training Details:Utilizing a large dataset of 11,000 images sourced from hypnohub, this LoRA focuses on a variety of hypnosis-related themes.2xT4 on Kaggle, 30gb vram in total. ~6 hrs training time. "I forgot the number but it's big" steps.Used code from Hollowstrawberry's Google Colab repo, modified, and adapted to Kaggle by me. Kaggle repo (code is a complete mess), find the Training SDXL Lora on Kaggle.ipynb file and import it into Kaggle.All dim alpha values was set to 24.Optimiser: ProdigyScheduler: REXBatch size: 6 (3 per gpu)Grad. acc. steps: 1Additionally used IP noise = 0.1; Prodigy 1.1.2 "slice_p=11" arg; --debiased_estimation_lossDataset:The training dataset consists of ~11k images scraped from hypnohub, all centered around hypnosis and related concepts. For a full list of tags used in the dataset, you can refer to the Bottom of the Tag Guide section where you can find a Gist link.Dataset was scraped from hypnohub with this prefill filtering:-manip,-animated,-animated_gif,-voice_acted,-animated_eyes_only,-tagme,-multiple_girls,-multiple_subs,-multiple_boys,-multiple_doms,-multiple_views,-traditional,-sketch,-meme,-ai_artAll images was downscaled to 1024 (bigger side) pixels by lanczos method in XnConvert tool.My booru scraper script also gathering tags from site. After additional filtering (see below), further tagging with JoyTag (high confidence like 0.6) in the Dataset Helpers tool.Additionally some further filtering were applyed:"score:>x" tag, where x value is defined by me for every main tag.Aspect Ratio filtering: for every image were calculated Aspect Ratio value (for example value for AR 16x9 will be 1.77), and images that were too long or wide (with an AR value greater than 3) were filtered out.After filtering and tagging, tags in each .txt file was organized by this metod:[Concept tags] = always first in the file, sorted a-z.[Style tags] = not sorted, simply moving them after Concept tags.[Main tags] = [Concept tags] + [Style tags]. The number of main tags is calculated. This is our keep_tokens value.After these steps, .txt files looking like this: [Main tags], other tags. Then all pairs of .txt + .img were sorted by keep_tokens value. Finally, I got a folder structure whose name depended on the keep_token value (i.e. 1, 2, 3, etc.).Tagging Guide:To utilize Mesmera effectively, you can explore a range of styles and concepts. Here are list of the main tags I was focused on:Concepts:before_and_after: 325 <-- 50/50. Works better with the "sequence"/"comic"/"2koma" tags.tech_control: 268 <-- Very base and non-specific tag. Model is slightly confused about it. Consider trying with the "hypnotic_audio", "hypnotic_screen" and other tech related tags.kaa_eyes: 244 <-- Not working. Needs better tagging/dataset filtering.bimbofication: 239 <-- 50/50. Needs better tagging/dataset filtering.happy_trance: 235enemy_conversion: 202 <-- 70% of times didn't work properly. Needs more dataset filtering.confused: 189pendulum: 185brain_drain: 183 <-- DO NOT TRY, GORY BRAIN IMAGERY! Will replace with "iq_drain"hypnotized_hypnotist: 179 <-- Kinda working.zombie_walk: 176hypnotic_screen: 174corruption: 168 <-- Already known by Illustrious and it's finetunes. Better concept generation achieved.hypnotic_audio: 166unhappy_trance: 154 <-- Tested only once. Works, but needed further exploration.hypnotic_app: 147 <-- 50/50pet_play: 147 <-- Already known by Illustrious and it's finetunes. Better concept generation achieved.shrunken_irises: 127 <-- 50/50visor: 125mantra: 113 <-- Works best with the "text" tag.standing_at_attention: 101brain_injection: 97 <-- DO NOT TRY, GORY BRAIN IMAGERY! Will replace with either "ear_penetration" or "mindfuck".asphyxiation: 85 <-- Not tested, sorry ..self_hypnosis, pendulum: 84 <-- Not working yet. Not enough representation in dataset.pov_sub: 84 <-- Not working yet. Not enough representation in dataset.harem_outfit: 81haigure: 81 <-- Not working yet. Not enough representation in dataset.hypnotic_tentacle: 56 <-- 50/50. Tentacle imagery and hypnotized person, but not hypnotic tentacle.dronification: 39 <-- Not working yet. Not enough representation in dataset.Styles:sleepymaid: 497jimryu: 338darkhatboy: 278 <-- Slightly noisy outputs given the original artist's style.hadant: 269oo_sebastian_oo: 220mythkaz: 194nexus_light: 193etlabsotwe: 177zko: 165 <-- Needs dataset filtering.eroborne: 164brellom: 160 <-- Not recognized. Needs dataset filtering/better tagging.onefeefoor: 159shishikasama: 149 <-- Works better with the "sketch" tag.katsiika: 146polmanning: 144faetomi: 143apopop: 138mahoumonsterart: 135smeef: 134djuuicebox: 134brushie_art: 134 <-- Not recognized. Needs dataset filtering/better tagging.zephyrgales: 133porniky: 128dochaunt: 123enetheligthingdancer: 122myuk: 121 <-- Basic anime artist. Nothing to learn for Illustrious. Consider slightly working.colorrings: 117 <-- Not recognized. I think that's because model "thinking" it's bad quality.medrifogmatio: 116crystal(zko): 114 <-- Character. Model didn't fully remembered her appearance. For example, no two color gradient on hair, only blonde.cursedrooks: 110erika_(er-ikaa): 109 <-- Character. Model didn't fully remembered her appearance. She simply resembles basic features.efalabrino: 108horiizyn: 108psyfly: 107artofadam: 106idpet: 1064headboiii: 104eshie: 102b-ginga: 102batta18th: 102kassidy_(medrifogmatio): 99 <-- Character. Working fully.ghostec: 99 <-- Kinda recognized. Needs dataset filtering/better tagging.pstash: 98 <-- Not recognized. Needs dataset filtering/better tagging.gmun: 96 <-- Basic anime artist. Nothing to learn for Illustrious. Consider slightly working.tomo86: 96abarus: 92vahn_yourdoom: 92supercasket: 91 <--3d artist. Significant image quality boost.httpwwwcom: 90keeper_of_pots: 90 <-- Not recognized. Needs dataset filtering/better tagging.yumiiart: 88erocoffee: 88lapislazuliart: 85electrickronos: 84ryuugu: 83wrenzephyr2: 83foolycooly: 82 <-- Significant image quality boost.zombi62: 79zorro-zero: 78davidthewolfx10: 73 <-- Much better works with furry characters.zelhypno: 72 <-- Not enough representation. Needs dataset filtering/better tagging.maozi_dan: 71maynara: 71 <-- Much better works with furry characters.reliusmax: 70jostony24k0: 69rosvo: 65malberrybush: 62 <-- Already known by Illustrious.yensh: 56sweetlittleneko: 56m4ns0n: 53 <-- Basic anime artist. Nothing to learn for Illustrious. Consider slightly working.kronobas28: 52nettleseeds: 504five1: 50 <-- Slightly noisy outputs given the original artist's style.cavitees: 49syas-nomis: 48shozaya: 47zelamir: 46mrs.erickson(zko): 41 <-- Character. Didn't test much, but i think it will work.konaloid: 40harvestman_here: 38 <-- Not enough representation.rnslivr: 38 <-- Not enough representation.shieol1: 37sleepy_gimp: 36 <-- Not working. Will replace with "sl33pyg1mp".gerph: 36blueparikeet: 36deepspaceart: 34magukappu: 34the_iron_mountain: 32drevod: 32strwbrrychz: 31glatu: 31cuddlesword: 29borvar: 29terasu_mc: 29 <-- Basic anime artist. Nothing to learn for Illustrious. Consider slightly working.konno_tohiro: 29hmage: 29kibazoku: 28singlesalt: 28 <-- Not enough representation.orphan2: 28hima_nsfw: 27 <-- Not working. Heavy bias to nsfw generation because of "_nsfw". Will replace with "himansfw".chien_vietnam: 27alittleshyart: 26a_singular_fish: 24slugbox: 20hypnoskeleton: 20 <-- Not enough representation. Also skeleton concept bleeding. Will replace with "hypnosk3l3t0n".rind_rin00: 20nez-box: 19thiccwithaq: 18 <-- Already known by Illustrious.sleepy_axolotl: 17 <-- Not working. Will replace with "sl33pyax0l0tl".ozoneserpent: 17cigarcat: 15 <-- Not enough representation.Full list of dataset tags can be founded in this Gist .txt file.Future plans for Mesmera version 2 (or 1.5, who knows):Will clean up dataset, better tagging;Less strict "score:>x" x value, generally even bigger dataset. Aim for 20k images.Fix main tags representation - try to balance the dataset, clean big tags, and hope that subset repeats do it's magic.Add these tags to main tags (list isn't separated by styles and concepts):progress_indicatoraltered_perceptionmicrochiprobotizationfembotameerashourdrawsantennahypnotic_accessoryhypnotic_musichypnotic_eyeshypnotic_breastshypnotic_gashypnotic_lighthypnotic_beamhypnotic_feethypnotic_penisfemdommaledommalesubfemsubspiral_eyesbodysuitlatexsee-throughclothed_exposureelectricityeye_rollglowing_eyescontrol_indicatorglowhornevil_smilenetorareauragameplay_mechanicsdefeatedtransformationsequenceempty_eyesbarcodecrotch_tattoobreast_expansionlip_expansionhentai_man