Mesmera

Copied from the page: Training Details: Utilizing a large dataset of 22,000 images sourced from hypnohub, this LoRA focuses on a variety of hypnosis-related themes. 2xT4 on Kaggle, 30gb vram in total. ~15 hrs training time, 8 epochs. "I forgot the number but it's B I G" 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: Prodigy Scheduler: REX Batch size: 6 (3 per gpu) Grad. acc. steps: 1 Additionally used IP noise = 0.1; Prodigy 1.1.2 "slice_p=11" arg; --debiased_estimation_loss Dataset: The training dataset consists of ~20k 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,-traditional,-sketch,-meme,-ai_art,-caption,-caption_only,-fascinum,-3d_custom_girl All 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: Meta: femsub: 700 <-- female submissive. use it to specify on whom chosen hypnotic effect will be applied before_and_after: 676 pov_sub: 557 <-- Point of view of submissive character sequence: 539 <-- can be used either as sequence of transformations of given character or just comic-style sequence see-through: 468 <-- contextually aware, i think... malesub: 426 <-- male submissive. same as femsub maledom: 309 <-- male dominant. use it to specify who is applying effects on fem/male sub. gameplay_mechanics: 289 <-- idk what is that, but it should help with some interfaces or level bars shit i guess femdom: 270 <-- female dominant. same as maledom pet_play: 222 Concepts: kaa_eyes: 700 happy_trance: 698 bimbofication: 683 hypnotized_hypnotist: 680 transformation: 553 <-- random transformations in training data, can help can make things worse breast_expansion: 512 <-- warying breast size, depends on artist style hypnotic_audio: 493 haigure: 480 dronification: 460 <-- warying results, general non-specific tag evil_smile: 444 hypnotic_screen: 435 bodysuit: 417 hypnotic_app: 415 tech_control: 402 <-- warying results, general non-specific tag control_indicator: 400 robotization: 365 clothed_exposure: 359 visor: 355 progress_indicator: 331 hypnotic_accessory: 329 lip_expansion: 326 zombie_walk: 316 hypnotic_music: 311 pendulum: 311 mantra: 304 altered_perception: 300 <-- use with "thought bubble" tag enemy_conversion: 291 hypnotic_eyes: 279 hypnotic_gas: 271 charm_(spell): 271 empty_eyes: 270 hypnotic_breasts: 269 electricity: 268 barcode: 266 hypnotic_penis: 262 hypnotic_light: 259 confused: 259 corruption: 242 chicken_pose: 235 spiral_eyes: 228 unhappy_trance: 212 shrunken_irises: 203 glowing_eyes: 201 hypnotic_beam: 184 standing_at_attention: 182 latex: 178 self_hypnosis: 176 housewife: 167 stage_hypnosis: 160 aura: 160 hypnotic_tentacle: 154 eye_roll: 152 hypnotic_ass: 139 harem_outfit: 132 hypnotic_feet: 129 asphyxiation: 103 Characters: erika_(er-ikaa): 510 kaa: 315 crystal_(zko): 226 hypno-tan: 162 kassidy_(medrifogmatio): 134 mrs._erickson_(zko): 59 Styles: sleepymaid: 624 jimryu: 512 darkhatboy: 345 oo_sebastian_oo: 331 hadant: 308 mythkaz: 267 nexus_light: 253 lairreverenteboladepelos: 243 etlabsotwe: 234 polmanning: 223 zko: 214 brellom: 200 katsiika: 199 onefeefoor: 196 eroborne: 196 zephyrgales: 177 mahoumonsterart: 177 apopop: 176 shishikasama: 174 smeef: 171 faetomi: 170 dochaunt: 167 porniky: 165 djuuicebox: 164 enetheligthingdancer: 160 ryuugu: 158 brushie_art: 152 efalabrino: 151 eshie: 151 glowhorn: 151 cursedrooks: 150 myuk: 150 psyfly: 148 4headboiii: 148 b-ginga: 144 detritus: 143 medrifogmatio: 137 artofadam: 127 batta18th: 126 ghostec: 126 gmun: 125 keeper_of_pots: 125 idpet: 123 horiizyn: 122 pstash: 122 <-- not working tomo86: 119 supercasket: 118 ameerashourdraws: 115 tunberuku: 113 <-- 50/50, dirty dataset httpwwwcom: 112 electrickronos: 112 erocoffee: 109 abarus: 109 yumiiart: 108 wrenzephyr2: 108 foolycooly: 108 lapislazuliart: 105 zelhypno: 105 <-- 50/50, why? idk :( supersatanson: 103 vahn_yourdoom: 101 zorro-zero: 100 yensh: 97 davidthewolfx10: 94 zombi62: 94 maozi_dan: 89 jostony24k0: 88 rosvo: 83 alerith: 82 reliusmax: 81 m4ns0n: 81 konaloid: 80 maynara: 80 sakurarose12: 79 sweetlittleneko: 78 malberrybush: 74 harvestman_here: 70 kronobas28: 68 orphan2: 68 syas-nomis: 63 zelamir: 62 nettleseeds: 61 4five1: 61 the_iron_mountain: 56 kibazoku: 55 cavitees: 55 sl33pyg1mp: 55 shozaya: 54 gerph: 50 blueparikeet: 49 drevod: 48 shieol1: 46 hmage: 45 konno_tohiro: 45 rnslivr: 44 deepspaceart: 42 singlesalt: 40 magukappu: 39 borvar: 38 glatu: 38 cuddlesword: 36 nez-box: 36 terasu_mc: 36 strwbrrychz: 35 thiccwithaq: 35 a_singular_fish: 33 hentai_man: 32 chien_vietnam: 32 alittleshyart: 30 rind_rin00: 28 slugbox: 24 Full list of tags in the Mesmera's training dataset can be found in this Gist. Mesmera v1 Gist links: Tagging guide: Gist. Full list of dataset tags: Gist. Roadmap (kinda): So... This LoRA is mostly working, but I'm still not satisfied with several quality aspects (for example, it was trained at 512 resolution). I will likely release Mesmera v2.5 this year, though I don't know exactly when. I feel slightly burnt out by this project, and I'm not going to push myself into it any further. I will focus on small-effort LoRAs for now. (27.03.25) After I release Mesmera v2.5 for Illustrious 0.1, I will likely follow this training plan: NoobAI Eps NoobAI V-pred Pony v6 And after Pony I will mostly handling user requests: Training for other models on the SDXL architecture: IL 1.1, IL 2.0/3.0 (hopefully), RouWei, Hassaku, WAI, etc. I will also try to train for Flux, Lumina, SD1.5, SD3.5, and Sana. Additionally, I will try to experiment with realistic style fine-tunes (Illustrious, Pony).