ponyxl_loras_n_stuff.md · hhks/PonyXL_Notes

#PonyXL Notes

[TOC2]

##Get PonyXL here https://civitai.com/models/257749?modelVersionId=290640

Anons have been experimenting with merging PonyXL with other checkpoints/loras, the following autism mixes have been popular, the DPO one has a much better default style out of the box, PonyXL LoRAs do still work on them but they may have a reduced strength. It seems these mixes generally have better hand anatomy although some complain that it gives it more of a 2.5D "aomslop" look.

https://civitai.com/models/288584?modelVersionId=324619

You can also experiment with the following lora on the base pony model, it's basically the difference between the autism_pony mix and base pony extracted into a lora, this way you can play with how strong you want the "autism" effect, this might allow you to preserve more of the appearance of the pony LoRAs you're using but also still get improved anatomy. You could also alternatively use an autism checkpoint and use this LoRA in the negative for a similar effect.

https://files.catbox.moe/kk4qjo.safetensors autismmix

There's a finetune of Pony with additional innate artist support now here: https://civitai.com/models/282341

An anon also created a LoRA extract of it if you want to experiment with it on base pony at varying strengths here: https://civitai.com/models/312010/4th-tail-lora-extract

##Reverse Engineered Hashed Tokens It seems like the obfuscated/hashed tokens in the model can often times be exposed through three letter strings (probably substrings of the full hash). There's only about 17k possible 3 letter combinations so it's possible to just reverse engineer the remapped tokens through brute force, an amazing anon went through the trouble of genning every 3 letter combination and noting which combinations seemed to affect the style. Here's his latest update:

All tokens I found: https://files.catbox.moe/41sbn0.txt Zip with all images (small) and tokens: https://files.catbox.moe/tqb4go.zip

Some more repetitive ones: aav to frw: https://files.catbox.moe/1z4efd.jpg fsp to klm: https://files.catbox.moe/jnjwqi.jpg kmq to ojn: https://files.catbox.moe/6qpxyi.jpg oka to rrg: https://files.catbox.moe/hvy7re.jpg rzj to xlv: https://files.catbox.moe/jrk1xl.jpg xmj to zzr: https://files.catbox.moe/882h83.jpg

There's an effort underway to identify the triggers.

###Hashed Styles/Characters

Update 3/21/2024 Here's a csv of the latest version of a collaborative spreadsheet with some identified styles. https://files.catbox.moe/j7g22l.csv

##List of Ponyxl LoRAs from /h/ NEW NOTE 1/24/2024: I've noticed that some LoRA bakers are using different negatives when testing their LoRAs, which can lead to discrepancies when using them. I've mostly noticed that some LoRAs have a significantly reduced effect if you don't have source_pony or some combination of the source_* tags other than source anime in the negatives. This is unfortunate but a byproduct of not having a standardized process for creating the LoRAs. There's another section down below that has more info about similar issues with PonyXL that's worth reading.

I made a powershell script to try and pull down the LoRAs on this rentry, it will store the last downloaded url for a lora in a history file in case a new or improved one replaces it so you can reuse the script without downloading everything again. You call the script and pass in a destination folder to store the LoRAs. Obviously run at your own risk, I can't prove to you I'm not an asshole or terrible programmer. You'll probably get throttled if you download everything in one go so run it overnight or something. Some of these LoRAs are huge, the whole folder with them all downloaded is ~24 gb.

UPDATE 1/21/2024 I added a basic management interface so it doesn't just pull down everything. You can pick and choose which loras you want. UPDATE 1/22/2024 Added some logic to keep old versions of LoRAs around in case one is updated but you still want to keep the old version around as well.

https://files.catbox.moe/2kiajb.ps1

###Random https://files.catbox.moe/kwo8w4.safetensors sayo_no_uta a horror game from 2003, you would have to look at the lora metadata to know what to prompt, there's flesh on the walls and stuff, fun to mess with https://files.catbox.moe/d9k7nf.safetensors slay_the_princess a visual novel that came out last year that has a sketchy art style, supports 5 of the game's characters: adversary, specter, nightmare, tower, and witch you have to prompt for their features to get them to show up, this file has the characters with their associated tags: https://files.catbox.moe/uprfne.txt, here's some example gens: https://files.catbox.moe/32eskz.png, https://files.catbox.moe/c87hl7.png, https://files.catbox.moe/kp5hbk.png, https://files.catbox.moe/jta8v1.png, https://files.catbox.moe/jsz1rp.png https://files.catbox.moe/te26ne.safetensors braces activation trigger is braces, some colors were tagged, rainbow braces, red braces, yellow braces, pink, braces, you pretty much HAVE to inpaint because the target area is typically going to be quite small in the image, could also try adetailer lips model https://files.catbox.moe/7723pp.safetensors doll a toy doll, activation trigger is doll, based on a set of photos so you'll have to use a style lora or some style triggers to make the image not a photo, works best with artists that have a more realistic style https://files.catbox.moe/tw9t1s.safetensors k-pop-stars a lora trained on kpop stars, don't think there's any activation triggers, I resized the original from pixel drain here: https://pixeldrain com/u/uku8ptP1, the anon has other ones at different step counts that might be better: https://pixeldrain com/u/5fuXqJp2 (4000 steps), https://pixeldrain com/u/wv6Vjq3v (2000 steps)

###Concepts https://files.catbox.moe/f6r1nl.safetensors shared_clothes main trigger is SharedClothes, example gen: https://files.catbox.moe/oddiam.png https://files.catbox.moe/7wpbsq.safetensors rogue_cosplay main trigger is RougeCosplay, dresses characters in rogue from the sonic the hedgeheg franchise, helper triggers are leotard, bodysuit, elbow gloves, thighboots and heart bustier, ears might bleed so it helps to put ears in the negative, example gen: https://files.catbox.moe/eeckj0.png

These are from the below mega, a couple were resized from their original size to get them under the 200mb limit https://files.catbox.moe/hfpslw.safetensors empty_eyes_drooling example gen: https://files.catbox.moe/p7aido.png https://files.catbox.moe/9qixqa.safetensors glowing_eyes activation trigger is glowing eyes https://files.catbox.moe/yz5c9g.safetensors quadruple_amputee activation trigger is quadruple amputee https://files.catbox.moe/tlt57h.safetensors ugly_bastard prompt stuff like ugly_male, ugly_bastard, old_male https://files.catbox.moe/ois9av.safetensors oral_cumshot_xray the activation triggers are cum in stomach, stomach (organ), cum in throat, lazy example gen: https://files.catbox.moe/h2j2dl.png

Resized ones https://files.catbox.moe/z71ic0.safetensors sex_machine activation trigger is sex machine https://files.catbox.moe/mxbbg2.safetensors stasis_tank activation trigger is stasis tank

These won't download automatically since they're on mega, most are in the above list, however anons sometimes add stuff to these folders that I might miss. https://mega.nz/folder/v2ZnADCR#hAELnn6ieCjsggLdG4quIw empty_eyes_drooling, glowing_eyes, quadruple_amputee, sex_machine, stasis_tank

###Artists/Styles These don't require activation triggers unless otherwise stated. Some of these had been previously uploaded to litterbox because they were too large for catbox. Unfortunately the links have expired, I've replaced them with resized versions that are just a couple ranks smaller so hopefully there's no quality loss. These replacement LoRAs have resized on the end of their names. I removed the original ones so people who were lucky enough to download them won't have them overwritten.

I generated example images for each LoRA, they can be found at this rentry: https://rentry.org/ponyxl_lora_previews, I started running out of space on this one.

These are from the megas below, unmodified, if you want good example gens you can check in their respective megas. https://files.catbox.moe/9qqq1s.safetensors suga_hideo https://files.catbox.moe/cmezo8.safetensors morino831 https://files.catbox.moe/yomld8.safetensors 11eyes https://files.catbox.moe/5vwp3d.safetensors alphy https://files.catbox.moe/08r1c0.safetensors baldr_sky https://files.catbox.moe/aymmt5.safetensors rai_sakuranbo_sugar https://files.catbox.moe/qmu9t3.safetensors yume_miru_kusuri

These are resized version of some of the LoRAs in the below megas so they fit on catbox and will thus download automatically with the powershell script, I didn't test them after resizing so can't comment on quality loss. I've shied away from resizing them now because it does seem to affect the quality a fair bit.

Megauploads/Other rentries with files too large for catbox https://mega.nz/file/xHt0kAQQ#lyOUMAOfm5MP60uNhCdT87Eaw8R6jB-oMpHbF_XBxEM mena https://mega.nz/file/pbszzZBA#WH08IpwvquRcueGie40IznEDMnV-jIsl-vLBaVfFWXg nullmeta (kokaki mumose) https://mega.nz/folder/83gQXTLT#mrjDP3w_OkxR0ujuVliesw kezime https://files.catbox.moe/b4wk0u.zip Megami_Magazine flat anime style https://files.catbox.moe/zxvyki.7z nyantcha https://mega.nz/file/V9V13KaQ#oeIASCTJrcK7CMJpFLHyqBWLlm0sGkFnvG9in8vKs9A slush https://mega.nz/folder/83gQXTLT#mrjDP3w_OkxR0ujuVliesw/folder/xuRSSbDK shirow,onono_imoko,kezime https://mega.nz/folder/RG81USjC#0T7bFau2u6bDCSodJJWX4w nullmeta,non_(z-art),tsukumizu_yuu,error1980,nns https://mega.nz/folder/2uBXjCRY#lPq-GnLJ1izbiJF0Y1N6TQ cham22, cyberboi, dratvan, gorani, lewdcactus, mmmmm, mojihe, owner, poper, whomperfruit, woohyootyeehawtown, zumi, furau, USjWwjFZLUtdSoz, senkan_art, inputtwo, jh, raira, zero0zero, too many to count https://files.catbox.moe/c69pcl.zip JVTeemo https://mega.nz/file/o01XhDIC#8hukpUUeFCThQj2WPe1O0OT7LDWHGvGmFr_y7GVUhfI borscht https://mega.nz/folder/iZcDnbiC#Yrxa1yRPxVd9F4pQQKuYpw suga_hideo, morino831, 11eyes, alphy, baldr sky, rai_(sakuranbo sugar), yume_miru_kusuri https://mega.nz/folder/dTkFHAgD#i04TRgRJUOMu9cByDe24UQ amazuyu_tatsuki, csr, enoshima, garakuta-ya, hiromitsu, kagami, kamitani, korotsuke, ohara, piromizu, JK-BlushySpicy-YD-CSR-Redrop_mix https://mega.nz/folder/OoYWzR6L#psN69wnC2ljJ9OQS2FDHoQ/folder/W5AjzbLD Fishine, Kakure_Eria, Navigavi, opossummachine, Shibori_Kasu, Syrupmelty, Zankuro, Kiss-Shot Azerola-Orion Heart-Under-Blade, Shinobu Oshino, Utsuho Reuji https://mega.nz/folder/YudnhQpC#KxYS978EU9mlyDwmLbIuwA/folder/tmtzgTxK katou_jun https://mega.nz/folder/k3MwHTwC#BOiRdd3uDjTeztzxuM19Hg alexi, chirumakuro, okiami, oltlo, yue https://mega.nz/folder/0soGXArQ#IJQJROng3TlELfCooa8RMg coro fae, danimaru, gonzrez, ishikei, kazutake hazano, letdi, nonco, oouso, rhasta, rwing, tottotenero https://rentry.org/shitty_loras#styles-pony-diffusion-v6-xl made_in_abyss, moshimoshibe https://mega.nz/folder/1yRSXQCR#Ziu9Q9Is5BCTYn53tnionw akatsuki_giken, alkemanubism, annin, aoin, chabo_fuketsudan, collagen, go_robots, ishida_umi, kafun, kaomin, kenya_suzuki, kink_(tortoiseshell), lucient, m_k, marui, miche, nezumi, nnf, pantsu_ruipper, RA, shizuoka, simzarts, testsuwan_woopie https://mega.nz/folder/NegX3DDL#NX3YguN8GSrEbOXuN918BA inhoso https://mega.nz/folder/fdoG2QDQ#6ZS_K55fY3g4NottW2NJng thanuki https://mega.nz/folder/L7Q3XaQC#jg7_M_gCun-aQenQmLISIQ banpai_akira, cluseller, tangobat, toi8, bonnie, everyntrge, gbf splashes (I assume some sort of gran blue style?), kaptivate https://mega.nz/folder/TZElCCIL#OEutCsJPD0G5lXPXLGgfcQ/folder/qMUFgTrR 2b213, Aestheticc-meme, Avotoma3, Blurent, Colodraws, Dustbinrat, Healthyman, Ibuo, Inkterton-kun, Jlullaby, Kairunoburogu, Kevbot, Korotsuke, Lakilolom03, Leftlewdz, Lone, Mantis-x, Marui Maru, Mctamagomaki, Murata Renge, Namespace, Possum Machine, Randomboobguy, Roly, Sandcavern, Sano-br, Sarikyou, Sweetdynamite, Takato Kurosuke, Vammzu, Wamudraws, Woofycakes, Yume Yoroi, Akai Sashimi, Fluffydango, Jujunaught, Jun (Rojiura Jack), Kanabun, Kanden-ki, Mystical, Spring2013, Bunybunyboi, Captainjerkpants, Citrusdoodles, Cloudxmoe, Coffeelove68, Doppel, Eigaka, Forastero, Gats, Gloss, Grayimpact, Greatm8, Kabos, Kanikama, Krekkov, Lk149, Loliconder, Marthedog, Naga U, Neo Hajime, Nuclearwasabi, Oekakizuki, Peculiart, Rizdraws, Ryusei Hashida, Sokonasu, Suzuki Kenya, Tod0231, Tomu (Tomubobu), Torathi, Zako

###Characters

https://files.catbox.moe/1wdl00.safetensors character_momoka helps to prompt features like hair, clothes https://files.catbox.moe/f5ssgz.safetensors character_tachibana_arisu helps to prompt features like hair, clothes https://files.catbox.moe/9zcayd.safetensors character_matoba_risa helps to prompt features like hair, clothes https://files.catbox.moe/djnlks.safetensors character_amy_sorel example prompt image here: https://files.catbox.moe/qymyf8.png, trigger is amy sorel, outfits are tagged as sc4 and sc6 https://files.catbox.moe/xow8kx.safetensors character_maryjane Prompt is maryjane, though it works better if you prompt for other stuff as well. gradient hair, tentacle hair, head fins, fangs, long tongue, etc https://files.catbox.moe/ccbfrs.safetensors character_neopolitan_rwby activation seems to be neo_politan, example gen: https://files.catbox.moe/m77z3a.png https://files.catbox.moe/0fmlta.safetensors character_naoto_shirogane looks like Shirogane Naoto is the activation trigger, example prompt image: https://files.catbox.moe/hf88vs.png https://files.catbox.moe/vsky78.safetensors character_shinobu_jackobs example prompt image: https://files.catbox.moe/jale97.png, activation trigger is shinobu jackobs https://files.catbox.moe/z8j9aq.safetensors character_roll_exe https://files.catbox.moe/ifr6a5.safetensors character_mayl_sakurai activation trigger is maylsakurai, can also prompt for the following 'blue vest', 'green sleeves', 'pink skirt', 'black thighhighs', 'hair ornament' https://files.catbox.moe/e9ult9.safetensors character_scarlet_ff7 activation trigger is scarlet_(ff7), also helps to prompt asymmetrical hair, blonde_hair, red dress if you want her in her typical red dress, looks best with 3d in the prompt due to the limited amount of images out there, once you put 3d in the negatives it starts to look a lot less like her https://files.catbox.moe/mxnwc6.safetensors character_fern https://files.catbox.moe/qxdzd4.safetensors character_luna_platz outfit infers very well as long as it isn't manipulated too much, you might see issues with things like dress/skirt lift. This LoRA is very flexible with other outfits known by your base model. Main activation token: lunaplatz Outfit: long-sleeve dress, lapel pin, striped pantyhose Optional: blue pantyhose https://files.catbox.moe/zrudlt.safetensors character_harp_note has a complicated outfit, so for full control over what is inferred by the model, there are very unique tags, and in all be pretty lengthy, especially for ponyxl. Main activation token: harpnote Outfit: helmet, tinted eyewear, pink dress, pink dress straps, bodysuit under clothes, striped detached sleeves, gold bracelet, scarf, pink kneeboots You don't need all of these tokens to infer her nsfw, usually the activation token with 'helmet' and 'tinted eyewear' is enough. https://files.catbox.moe/j7p2f3.safetensors character_ame-chan Trigger token is ame-chan https://files.catbox.moe/h3udxn.safetensors character_neptune Trigger token is smolneptune Outfit: hooded jacket, white choker, d-pad hair ornaments, blue striped thighhighs, blue striped panties Can also closely replicate her frilled swimsuit with 'orange frilled bikini'. https://files.catbox.moe/80ioqv.safetensors character_nepgear Trigger token is nepugia Outfit: long-sleeve sailor dress, d-pad hair ornament, white choker, pink striped thighhighs, yellow neckerchief, pink striped panties, thigh strap https://files.catbox.moe/uveemt.safetensors character_hanabi Trigger phrase "hanabi hsr" to get her to gen, "fox mask" for her mask(s), "mask on head" to get the mask on her head, "black gloves" for her gloves https://files.catbox.moe/9sc9aj.safetensors character_hiragi_kagami example gens: https://files.catbox.moe/6j5etd.jpg, https://files.catbox.moe/22j10a.jpg, https://files.catbox.moe/vkuomy.jpg, https://files.catbox.moe/855i66.jpg https://files.catbox.moe/p8pc8o.safetensors character_takarada_rikka doesn't seem to have an activation trigger, just will make any girls you gen into her https://files.catbox.moe/0520b4.safetensors character_ashley_warioware activation trigger is ashely, looks like the images all had 1girl in the prompt as well https://files.catbox.moe/0r3b9z.safetensors character_beatrice_rezero doesn't seem to have a trigger for the character https://files.catbox.moe/xv8o6t.safetensors character_sparkle_honkai_star_rail looks like sparkle_(honkai: star rail) is the activation trigger, example gens: https://files.catbox.moe/fz15l8.png, https://files.catbox.moe/mvxmzo.png, https://files.catbox.moe/tpwdsm.png https://files.catbox.moe/26s9wl.safetensors character_purah activation trigger is purah. example gen: https://files.catbox.moe/8oh179.png https://files.catbox.moe/mugz6g.safetensors character_elegg_nikke activation trigger is elegg https://files.catbox.moe/ydfkfj.safetensors character_seele_starchasm_nyx activation trigger is seele vollerei (starchasm nyx), example gen: https://files.catbox.moe/oki6x0.png https://files.catbox.moe/0escz4.safetensors character_sindel activation trigger is sindel, helps to have multicolored_hair in the prompt to get the black streak in her hair to show up https://files.catbox.moe/pagyj3.safetensors character_agnes_tachyon activation trigger is agnes_tachyon_(umamusume) https://files.catbox.moe/ou15oe.safetensors character_rishiah_ivyred activation trigger is rishia ivyred, seemd like you need to prompt her features as well like green hair, green eyes, twin braids, very long hair, hair bow. example gen: https://files.catbox.moe/5t64vg.png https://files.catbox.moe/62g6g1.safetensors character_uni seems like un1 is the activation trigger, example gens: https://files.catbox.moe/3ijdtz.png, https://files.catbox.moe/rgndp3.png, https://files.catbox.moe/t306wb.png

Mega uploads with files too large for catbox https://mega.nz/folder/uuI2DCjA#JKGPF0uYT-1Zo0dE3wCPJQ sorceress_sellen https://mega.nz/folder/L7Q3XaQC#jg7_M_gCun-aQenQmLISIQ characters from nikke, hsr, yugioh, azure lane, blue archive, gran blue

##Innate artists/styles List of tags innate to ponyxl that change the style significantly, usually artists but sometimes other strange behavior some of these could just be getting activated by character names in the artist name or just misfires in the AI network ¯_(ツ)_/¯

asura_(asurauser)

mirai_hikari

cutesexyrobutts

bb_(baalbuddy)

dd_(ijigendd)

hara_(harayutaka)

hews

boris_(noborhys)

m-da_s-tarou

tsukishiro_saika

tsukudani_(coke-buta)

agawa_ryou

sumiyao_(amam)

neocoill

tomose_shunsaku

null_(nyanpyoun)

horosuke

chanta_(ayatakaoisii)

merunyaa

fumihiko_(fu_mihi_ko)

aaaa_(quad-a)

cle_masahiro

kasumi_(skchkko)

murakami_suigun

fumio_(rsqkr)

yomu_(sgt_epper)

homare_(fool's_art)

happoubi_jin

asamura_hiori

moisture_(chichi)

dandon_fuga

kase_daiki

taka_(taka_wo_kukuru)

ririko_(zhuoyandesailaer)

mo_(kireinamo)

akinbo_(hyouka_fuyou)

aztodio

kaisen_chuui

sousouman

bee_(deadflow)

sano_toshihide

mitsumi_misato

personal_ami

donguri_suzume

lambda_(kusowarota)

jack_dempa

lm_(legoman)

minakami_(flyingman555)

jellytits-7

amasora_taichi

em_(totsuzen_no_hakike)

kurokoshou_(emuburemu123)

niliu_chahui

kawakami_rokkaku

miyase_mahiro

momo_no_sukebe

minato_ojitan

joy_ride

osiimi

mosho

fuya_(tempupupu)

jabara_tornado

konno_tohiro

23_(real_xxiii)

kashu_(hizake)

mizuryu_kei

kairunoburogu

mizuumi_(bb)

nagase_haruhito

nottytiffy

akashio_(loli_ace)

liang_xing

vahn_yourdoom

nakamura_regura

aster_crowley

cian_yo

hekiga_(freelot)

tsukiyo_(skymint)

nyong_nyong

tony_guisado

tebi_(tbd11)

neoartcore

yoi_naosuke

damao_yu

masami_chie

from here will try to classify them better, too lazy to revisit past ones strong and accurate (may need strength tweaking)

tony_taka

dross

hua-j

sparrow

firolian

paloma_piquet

akano_murasaki

Kawamoto_Toshihiro

blade_(galaxist)

bittenhard

stealth-brock

incase

sunna_(nilsunna)

wagashi_(dagashiya)

sagiri

melkor_mancin

alp

elf_(stroll_in_the_woods)

tomu_(tomubobu)

kame_(kamepan44231)

cyancapsule

kelvin_hiu

neisan

pija_(pianiishimo)

koikatsu

mamimi

$source_anime, score_9, from behind, cowboy_shot, 2b $nier:automata$, ass_focus, leotard, hands on hips, looking back NEG: monochrome, greyscale$ mdf_an

strong and doesn't look anything like the artist imo but still possibly useful through varying strength and combining with other artists geo_(yukishitadou)

fadingz

sima_naoteng

nyororiso_(muyaa)

seems to induce a change but I could be hallucinating or it could be just due to the randomness of the model

nemunemu_(candy_paddle)

kurokaze_no_sora

kapuru_0410

iro_ame_(amewaagada)

hisakawa_chin

sunga2usagi

shirokuma_(nankyoku)

kokura_masashi

hinaki_(hinaki_0102)

haru_(hiyori-kohal)

hanamasa_ono

amano_kozue

ziro_(zirorong)

shiromoru_(yozakura_rety)

sakura_hanpen

nyonyonba_tarou

miyuki_(miyuki0529)

ichirino_minagi

shimashiro_itsuki

nijihashi_sora

mugichoko_(mugi_no_choko)

kamimiya

haneramu

nakamura_rohane

oimo_(imoyoukan)

mugi_(banban53)

mana_kakkowarai

last tried was yanyan_(shinken_gomi)

##Innate Characters I spent some time generating images of the characters on danbooru in descending order by post count, here's about 500 of the ones I saw that pony recognized.

https://files.catbox.moe/r29xho.txt

##Tricks with the model for more anime style There's some evidence that the model has a bit of a bias towards western style art, particularly when using the score tags (it is made by furries after all). It might be beneficial to tag your image sets with the score tags, the laziest way would be to tag every image with score_9, source_anime, this may affect the overall "quality" of images the lora generates since some of the knowledge in the model about "high quality" might be overwritten but it will look more like the artist's style.

You can also help nudge generations with the model more towards anime style by using "source_cartoon, source_furry, source_pony, sketch, painting, monochrome" in the negative prompt when generating images. Some of the artists with more subtle art styles tend to have more success with this in my experience. The bias is a bit less evident in the lower scores so if a LoRA has it's images tagged with score_9 a prompt like "source_anime, score_9, score_6_up, score_5_up, score_4_up" might get better results. Unfortunately this also makes your LoRA harder to use, you'd have to tell people to use it this way.

Other tricks tried were to include all the score tags in the LoRA's training set on every image which didn't have much success or only using the source_anime tag which also didn't seem to influence the LoRA's effectiveness much. I haven't tried actually using the score tags as intended since that takes more effort than I'm willing to put in for a LoRA for the time being but that may net the best results.

##Various Anon LoRA Training Settings ####Specific block alphas and dims I've been seeing some pretty good LoRAs that are making use of the following settings, I couldn't really tell what these settings are doing that's different from just setting the overall dims/alpha of the LoRA but I have seen some of the generated LoRAs and they do seem good, especially considering they're typically just 40 megs.

block_alphas=0.0625,0.0625,0.0625,0.0625,0.0625,0.0625,0.0625,0.0625,0.0625,0.0625,0.0625,0.0625,0.0625,0.0625,0.0625,0.0625,0.0625,0.0625,0.0625,0.0625,0.0625,0.0625,0.0625,0.0625,0.0625 block_dims=8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8

There's also been a couple anons playing with the multires settings like so, again I couldn't tell you what exactly this is doing, maybe adding some noise to the images while training to prevent burning up the network?

multires_noise_discount=0.3 multires_noise_iterations=6 or multires_noise_discount=0.5 multires_noise_iterations=8

Here are a couple shared json config files that have utilized these settings, these settings also reduce the effective learning rate significantly so you'll notice the unet learning rate is much higher than you might typically expect: https://files.catbox.moe/p1m50y.json https://files.catbox.moe/fgp24t.json https://files.catbox.moe/s75mc6.json

Typically with these settings anons are trying to hit 1.5k-3k steps, so adjust your epoch count based on image count accordingly.

I have incorporated the multires_noise_discount and multires_noise_iterations into my own trainings and it seems to improve the quality, although to be honest I haven't really been able to hit the quality I like at just 2k steps and am still targeting more like 5k+, I'm also still mostly using prodigy as my go-to optimizer.

####Custom scheduler Another anon that has had some good LoRAs has gone a different route with a custom consine annealing scheduler:

https://files.catbox.moe/ua9jua.toml

####Prodigy and multires

There's another anon that's been using prodigy with the above multires settings which is what I've been doing lately, seems better than the old settings I was using.

https://files.catbox.moe/m861nn.json

####My Current Settings My current settings are pretty similar to the prodigy one above, a couple times I've had a hard time getting prodigy to get that last bit of distance to really recreate the artist and have used AdamW in those cases, which takes a bit more tweaking with the learning rates. Usually I'll resize the resulting LoRA down to 16ish dims (for a style) and make sure it still looks good.

--train_data_dir = "$IMAGES_PATH$" --output_dir = "$OUTPUT_FOLDER$" --output_name = "$LORA_NAME$" --pretrained_model_name_or_path = "$PATH_TO_PONYXL_CHECKPOINT$" --max_train_epochs = $NUMBER_EPOCHS$ --train_batch_size = 8 --resolution = "1024,1024" --save_every_n_epochs = 1 --save_last_n_epochs = 999 --learning_rate = 1 --lr_scheduler = cosine --network_dim = 64 --network_alpha = 32 --seed = 1337 --keep_tokens = 0 --gradient_checkpointing --max_data_loader_n_workers = 16 --mixed_precision = bf16 --save_precision = bf16 --network_module = lycoris.kohya --network_args "preset = unet-transformer-only" "conv_dim = 64" "conv_alpha = 32" "algo = locon" --optimizer_type = "prodigy" --optimizer_args "decouple = True" "weight_decay = 0.01" "d_coef = 0.8" "use_bias_correction = True" "safeguard_warmup = False" "betas = 0.9,0.99" --multires_noise_iterations 6 --multires_noise_discount 0.3 --caption_extension = ".txt" --prior_loss_weight = 1 --cache_latents --cache_latents_to_disk --xformers --save_model_as = safetensors --enable_bucket --min_bucket_reso = 512 --max_bucket_reso = 2048 --bucket_reso_steps = 128