Hassaku aims to be a anime model with a bright and distinct anime style.
My Discord for everything related to anime models and art.
If you'd like to support my work, you can do so at: SubscribeStar, every bit helps and is truly appreciated!
____________________________________________________________
Supporters:
Thanks to my supporters tamashiicolle, pttcot, SETI and Kodokuna
____________________________________________________________
Using the model:
The model was trained primarily on images with minimal disruptive elements, such as floating text, logos, speech bubbles, or signatures. If any of these elements appear in an image, please include "signature" in your negative prompt.
Danbooru Metadata and franchise tags are not used in training, so please avoid using them in prompts. Metadata such as "highres" and franchise tags like "re:zero kara hajimeru isekai seikatsu" are not used in training.
A CFG scale between 3 and 7 is recommended. For best results, use CFG 6 with the Euler a sampler.
Refer to the example images, to understand, how the model should be used.
The model is designed to be simple and straightforward to prompt. Use the following tag order:
Number of characters
Character Tags
All remaining tags
Example: 1girl, rem \(re:zero\), standing, masterpiece, upper body
Here are some resolution options for SDXL:
1536 x 640
1344 x 768
1216 x 832
1152 x 896
1024 x 1024
896 x 1152
832 x 1216 (most recommended)
768 x 1344
640 x 1536
Lora use:
NoobAIs Loras working for the most part better in compare to Illustrious loras
Versions 2 and 3 is less compatible to most loras trained on Illustrious or Noob, because it got further trained away of both. Style A is still the most compatible for it.
______________________________________________________
Version and License info:
Below V1 merges ANIMAGINE XL 3.0
Version V1 use Illustrious-XL & WAI-NSFW-illustrious-SDXL with additional training
V2 is trained on its own and don't include any extra merge. Base was V1, it is trained to include newer or missing characters. It was also used for further training tests
V3 is a merge of Illustrious-XL and V2.2, to fix some issues and was also trained to include newer or missing characters
All Models using the Fair AI Public License 1.0-SD license.
Description
Model was trained on more actual data, but only on a small dataset of 50k images (mostly mihoyo games). With a month of experimenting and a hard reset inbetween, because of better training settings. New characters are relative weak trained in and the style merge mostly erased details of them, in which mostly the clothing suffers. It is more or less a prototype to not make the same mistakes of V2.1, that resulted in sampler problems.
better (all in compare to 2.1fix)
- Less orange looking images
- Style is overall more even torwards anime
- Model is a bit less restrictive and more expressive
worse for some
- more anime looking, which also makes backgrounds less detailed and less cohesive
- less mature looking faces
- Can be darker on some concepts (also more shadows on faces)
- Pupils are more filled out (less circle pupils)
worse
- hands/fingers
- text (model goes more towards japanese writing, when it wants to generate text, which is disjointed)
Can't say
- Loras should work better, because of a better inner style merge, with much less artifacts. But i can see that a strong style can backup loras to work better. It mostly depends of the loras itself.
2.1fix have a more 2.5d feel to it, like more details on bodys (like showing rips more often). Which also translate to backgrounds. If you like that, you can still use 2.1fix further.
If you question why i can't stay on the same qualitys of 2.1fix, its because i can't be more "reckless" in my mergin, without actual hurting the new knowledge in the model. Mergin another model in, erase every training I done on it, because all other models knowing less then my model. While every longer training on it, resets it to be like every other base model like noob or illustriuos, without style merge.
FAQ
Comments (198)
I came to IL from SD1.5 Koji, which checkpoint should I use to get similar results to Koji? What's the differences between Hassaku, Amanatsu, and Sudachi?
I tried some other IL checkpoints but they all seem to produce cartoony results, like they're too clean and feel like western arts instead of Japanese. Is that what base IL doing?
After trying all three checkpoints, it seems like Amanatsu is what I need, I think? It's producing images with more details, more noises and artifacts, so that it looks more like real drawings instead of clean computer arts that base IL do.
Fantastic model 💖! What clip skip do you suggest we use with this?
I don't get the hype for Illustrious. EDIT: NVM I'm a dumbass. I tried using my same Pony prompts and yeah... bad idea. Illustrious is actually really good at using natural language to describe what you want the scene to look like.
Actually ☝🤓, no. Illustrious isn't better when using natural language, that's FLUX.
so in short, 2.2 has slight downsides compared to 2.1fix?
In some aspects yes, since I making models, i also list downsides on my updates and i don't want to stop that, so everybody can decide self, what they want to use👍
@Ikena right, understood, i'll personally stick to 2.1fix due to there being less listed downsides, maybe when a more fixed version of 2.2 or higher is out, i then might try it, until then, 2.1fix is good enough for me.
Is 2.2 good enough compared to 1.3 Style A? I didn't switch to 2.1 fix, because it has many problems compared to 1.3 Style A.
Can you say what problems 2.1fix have in compare to Style A?
Whats new in v 2.2?
shame need to wait 7 days for test it my self then i read the discription of 2.2 and i think wait a minute hands worse sometimes darker? what? tested 2.1 + 2.1fix i am not sure what it fixes exactly out of hands but 2.1 has on resolution 1024x1024 so much problems that i get cropped images like focus on the body and it does what it want if u write from side or something sometimes i get a img from behind and if i write not sky or clouds and only city the whole city is bright like alot of fog. so yeaah i apreciate the time that u spend in but i like the adult style taste of 2.1 and the drawing but what i hate on this is getting weird poses or cropped images ( the fog is ok if u write more stuff for the background ) if u think to work again on a 2.2 fix or maybe a better 2.1 then keep the adult style and watch on resolution 1024x1024 that u dont get cropped images , maybe it was a incomatible model that u merged in ur trained model so far i readed first the whole stuff and then i tested it :/ then i need to wait for a version thats like 2.1 drawing but not with the problems maybe it happens or maybe not i dont know its ur choice as creator but wish u good luck for the next versions that it works better greetings
After all the update still love the 1.3 the most 😂
v1.3a supremacy still strong and unbeatable 💪
@HahaAigoBrr full body is more like a opposite to close-up, did you use prompts that describe facial features like eye colors, so the model knows you mean a face with it?
@Ikena Yeah I described facial features, body features, all that, but still get feet a lot. I also realised it doesn't really know what a 'handjob' is and I get a lot of blowjobs when the subject is kneeling. Also this by no means is hate to the model, I use it almost EXCLUSIVELY now because of how much I love it (though I use styleB)
nice
Undoubtedly the best illustrious checkpoint. Even better than ntrmixillustrious
The best ai model I have ever used
试了一下2.2，依然很棒，但感觉绘制特定人物特征的能力与2.1fix相比好像略有下降。
I tried 2.2, it's still great, but I feel like the ability to draw specific character features has dropped slightly compared to 2.1fix……maybe
When are you bringing the new 2.2 model to ShakkerAI? By the way, this new version is amazing — by far the best Illustrious model so far!
When its out of early access, i upload it also to other platforms👍
Is it possible to gen Cantarella and Trigger with this model w/o using lora now? O:
Yess
Style B is very unique and beautiful.
Hope they continue to develop.
Style B is here https://civitai.com/models/1325426/amanatsu-illustrious
@Ikena Ohh Thank you!
perfect
From which website does it uses its tags from? Or rather how do i use tags for the prompts?
is that possible to train next version with more data which contains "scenes", like interactions, more charachters, or is that too expensive for that type of finetunes? I just dont know
yes, new characters are my main focus now. But new concepts or scenes or even artstyles of artist are possible and are also being trained in with the new characters.
@Ikena nice.
@Ikena Hey, a request for characters from metafor refantazio and unicorn overlord to be included in
@zanvanser Sure, adding them to the dataset, i already train on a new dataset, so i can't promise that the characters will work as great in compare to the other characters. Especially for unicorn overlord, with only around 15 images for some characters. That is not enough. But rosalinde (unicorn overlord) and scarlett (unicorn overlord) will probably work (the ones with more then 100 images)
Comparing to 2.1fix, it seems I have more extra limbs/heads issue with 2.2 after testing for a few days. The same prompts often do not result in the extra limbs in 2.1f comparing to 2.2.
Thanks for the feedback, is it tested on specific concepts or on a wider field of themes?
@Ikena Thanks for the prompt reply, I am new to image generation so I am afraid I can't judge if my scope is wide. My prompt was not too specific in terms of gesture, sometimes it's like "laying down facing the viewer" or more often there weren't any prompt for specific pose. About 3/10 times, images might come out with an extra arm or leg, sometimes maybe an extra head with V22, and with V21f it maybe more like 1/10 or fewer. (these are very rough numbers).
This also seems to happen more if I used a higher resolution than suggested, such as 1920x1080.
Overall, it's a small comment from a quick observation, and is not so problematic and can be prompted away with a few shots. I am very happy with the results so far. Thanks for the great work!
Just tried a bit more, if I stay with 1024*1024, things are almost always fine. The moment I tried 1920*1080, I start to see extra limbs/heads. So it appears staying within the suggested resolution is critical.
@sandbo Staying in the right resolution on what the model is trained, prevent it more to happen. I also see that you use "laying down facing the viewer", so you don't use tag based prompting? A little suggestion, maybe try it out with tags of danbooru, it probably also helps to get what you want👍
@Ikena Thanks for the tips, I learnt a lot lately!
peak
Good
What VAE do I use with this everything just comes out looking like its melting
The model includes a vae, do you use comfy? do you use loras? because it should not look like its melting
Same for me in A1111, when i try to do png info -> upload a photo -> send to txt2img --> generate
terminal shows a message like "couldn't find VAE named sdxl_vaeFix.safetensors --> using none Instead"
@Nozaki_kun Thats a setting issue, because that error shows up, if it wants to use a external vae and not the vae in the model itself. Do you set a vae in the settings to use?
@Ikena nope, in VAE folder there's nothing
@Nozaki_kun settings -> vae, is a vae selected?
@Ikena nope, it is on automatic
Choose "none" instead of "automatic"
@EbenezerDanglewood same :(
Using v2.2, I've had almost no problems, but one thing that often happens is that the color of the vagina ends up being the same color as the lipstick or a very strong pink, it could be a lack of experience, as I'm new to Illustrious
Can you give the promt please?
@Ikena sure, I'll send you a private message
Doing god's work, @tomatosad
You, too, @Ikena
10/10
hmm more accurate to prompt, but seems to have more style interference than v12, which was much more cleaner. in pictures with multiple views of a character it also adds a bunch of garbled mess faces. meh.
Next version 2.3 new characters and outfits for known characters currently in training (some are already in 2.2): https://drive.google.com/file/d/1X5HYJq4yyfmRX0NIhtUKIILofnV_eQbG/view?usp=sharing
You can make suggestions, what i should also add to it. Can't promise that i add it, Training is expensive and i mostly adding new characters in with at least 100 images.
isaki_riona
mizumiya_su
kikirara_vivi
koganei_niko
rindo_chihaya
100 to 320 images on danbooru
@Nyaa314 thanks! all are new around the 6 month mark, where i'm not sure, if they are included already or not, but it seems to be not included in previous trainings. Adding all to the current set👍
Pokemon trainers.. May, Serena, Dawn.. in Pokémon Contest outfits.
@CelestialDawn if those contest outfits are new and as tags on danbooru present, i can train it in. The characters you named are already trained in. As far for Pokemon, I train taunie_(pokemon) and hex_maniac_(pokemon_legends:_z-a) currently newly in
@Ikena may_(contest)_(pokemon) is a tag, I think, for May's contest outfit. Outfits like that one would be nice to have baked into the model.. also outfits like; ballet, leotard, bodycon, magical girl, princess dress, etc. !
exusiai_the_new_covenant, mon3tr
@hectorq323964 thanks, mon3tr is included, but not in the list i linked here. mon3tr is not new, reason why its not in the list, only her girl form is new, but on a old tag... i included her myself. exusiai_the_new_covenant don't have 100 images, but i add her nevertheless
may_(contest)_(pokemon) and miranda_(nikke)
anmi and avian_romance
@CelestialDawn Please characters/desigs that are new with at least 100 images. I only add exusiai_the_new_covenant with less then 100 images, because that design is only 6d old. You are naming characters/desigs that are already up to 6years old with less then 100 images.
@Ikena I'm curious, does this model only work with danbooru? Who even uses that site anymore? The images on there are mostly terrible and nobody I know of has used it in years..
wilhelmina_(brown_dust),
loen_(brown_dust_2),
morpeah_(brown_dust),
eclipse_(brown_dust_2),
teresse_(brown_dust),
liatris_(brown_dust),
So with danbooru it's also not possible to train things like tattoo choker, chock collar, etc because that website barely has any images on it? I will try to find a lora maybe
@kiragira24 I'm not familliar with brown dust, Illustrious can't gen any characters of it? with 1,3k images, i can simply include "brown_dust_(series)" entirely
@CelestialDawn yes, like every other anime model. It is the biggest tag based image database of the world. My model also included e621 and gelbooru as source.
@CelestialDawn for such special niche tags, no, but there are so many other collar variants https://danbooru.donmai.us/wiki_pages/collar?z=2 with in sum 200k images that you can use instead. Also for choker https://danbooru.donmai.us/wiki_pages/choker
@Ikena there are Brown Dust 2 offical arts even some erotic video are offical ingame scenes in pg13 game
https://gelbooru.com/index.php?page=post&s=list&tags=brown_dust_2+official_art+&pid=0
https://gelbooru.com/index.php?page=post&s=view&id=11727508&tags=brown_dust_2+official_art+
https://gelbooru.com/index.php?page=post&s=view&id=11447928&tags=brown_dust_2+official_art+
https://gelbooru.com/index.php?page=post&s=view&id=11447925&tags=brown_dust_2+official_art+
@kiragira24 thanks, but that was not my question... i looked up and i had some of the characters in my previous datasets. Some of the characters barely work on testing, so i include a bit of brown_dust characters in. But on lower priority
Some Characters from Tower of Fantasy like Nemesis, Voidpiercer, Greyfox, Meryl or Ironheart.
Tamaki_Kotatsu, Shiina_Mashiro
@Ikena gelbooru is nice, models that use that tend to be better
@Ikena are u gonna add the new ZZZ and Wuwa characters to it? O:
@RizzieLove yes, i already train them in, but its only recently with not many images, its not working right now right. But any character in the list here should work in the next version
@Ikena thank u for responding my dude :D
kaguya (blue archive) could be a good edition, she has about 300 images :)
Teacher, Google Drive is invalid. Are there any new links
@254872742 i have no idea why google removed it, but here is the newest list:
conch_(gundam), yuuki_sakuna_(1st_costume), vivian_banshee, curly_(pre-crash)_(mouthwashing), xiangli_yao_(wuthering_waves), michigami_nareko, perfect_start_(infinity_nikki), metatron_jeanne_(fate), shiiko_sugai, exusiai_the_new_covenant_(arknights), rio_(bodysuit)_(blue_archive), aoba_(blue_archive), mortis_(persona)_(bang_dream!), yi_xuan_(zenless_zone_zero), netherwing_(honkai:_star_rail), lapwing_(vrchat), cloud_strife_(midgard_seventh_high_school), aerith_gainsborough_(midgard_seventh_high_school), amane_konata, harmony_(pokemon), ouro_kronii_(maid), escoffier_(genshin_impact), pius_(aotu_world), xander_(aotu_world), bready_(nikke), mogador_(olfactory_nursery_call)_(azur_lane), fleurdelys_(wuthering_waves), franklin_(azur_lane), reed_the_flame_shadow_(summer_flower)_(arknights), klukai_(speed_star)_(girls'_frontline_2), ju_fufu, lind_(hotathino), xia_fei, white_haired_girl_(chainsaw_man), wahoo_(kancolle), jk_mob-chan_(minami_aomori), dante_alighieri_(fate), hex_maniac_(pokemon_legends:_z-a), taunie_(pokemon), psylocke_(vengeance), professor_niyaniya_(teacher)_(blue_archive), juri_(part-time)_(blue_archive), sena_(casual)_(blue_archive), lucky_lilac_(umamusume), chrono_genesis_(umamusume), curren_bouquetd'or_(umamusume), fenomeno_(umamusume), gran_alegria_(umamusume), almond_eye_(umamusume), tribios_(honkai:_star_rail), milltina_(vrchat), mori_calliope_(kawaiiope), recluse_(elden_ring), varesa_(genshin_impact), mechty_(girls'_frontline_2), wakaba_mutsumi_(doll), entelechia_(arknights), cantarella_(wuthering_waves), xiao_weiying, shirogane_noel_(battle_nun), houshou_marine_(loungewear), comoli_harcourt, xavier_olivette, kiana_kaslana_(ba-dum!_fiery_wishing_star), miyaboo_(zenless_zone_zero), asaboo_(zenless_zone_zero), pulchra_fellini, koseki_bijou_(casual), mococo_abyssgard_(casual), fuwawa_abyssgard_(casual), nerissa_ravencroft_(casual), challia_bull, mrs._snake_(fate), trigger_(zenless_zone_zero), anby_demara_(silver_soldier), ceras_yanagida_lilienfeld, mast_(romantic_maid)_(nikke), mafuyu_(vrchat), shu_(spring_feast)_(arknights), naplings_(nimi_nightmare), maliss_<p>_white_rabbit, trinnon_(honkai:_star_rail), trianne_(honkai:_star_rail), yu_(arknights), nimi_nightmare, astra_yao_(chandelier), shiori_novella_(casual), cartethyia_(wuthering_waves), ellen_joe_(on_campus), mori_calliope_(rock_bunny), oozora_subaru_(police), citlalin_(genshin_impact), itzpapa_(genshin_impact), crazy_mita_(miside), yumemizuki_mizuki, cure_wink, aokaze_nana, shigure_kokoro, cure_kyunkyun, pero_(wuthering_waves), sakura_uta_(precure), liang_yue_(reverse:1999), raora_panthera_(new_year), cure_idol, klukai_(girls'_frontline_2), hugo_vlad, laimu_(vtuber)_(cozy), jinhsi_(peach_blossom)_(wuthering_waves), j_(reverse:1999), rapi_(red_hood)_(nikke), momomiya_shizuku_(gaoyue_jixiang), shinano_(visions_of_fantasy)_(azur_lane), yvonne_(arknights), hu_tao_(cherries_snow-laden)_(genshin_impact), kind_mita_(miside), cool_mita_(miside), z11_(azur_lane), evelyn_chevalier, maika_(ogata), mita_(miside), astra_yao, kaya_rinha, lattice_dragonmaid, louhi_(fate), anyoji_hime_(winter_loungewear), abigail_williams_(santa)_(fate), houshou_marine_(ahoy!!_you're_all_pirates), shuji_ito, anyoji_hime_(summer_loungewear), kachimachi_kosuzu_(summer_loungewear), momose_ginko_(summer_loungewear), kisaki_(school_uniform)_(blue_archive), nyaan_(gundam_gquuuuuux), amate_yuzuriha, gquuuuuux, barcarola_(reverse:1999), maiden_(ice_rose)_(nikke), mr._scarletella, lan_yan_(genshin_impact), cyrene_(honkai:_star_rail), phainon_(honkai:_star_rail), cipher_(honkai:_star_rail), mydei_(honkai:_star_rail), hyacine_(honkai:_star_rail), anaxa_(honkai:_star_rail), tribbie_(honkai:_star_rail), castorice_(honkai:_star_rail), zani_(wuthering_waves), roccia_(wuthering_waves), brant_(wuthering_waves), phoebe_(wuthering_waves), mr._hood, mem_(honkai:_star_rail), aglaea_(honkai:_star_rail), the_herta_(honkai:_star_rail), imadake_dabuchi_tabemi, carlotta_(wuthering_waves), raidian_(arknights), heroine_(homicipher), shorekeeper_(wuthering_waves), mr._crawling, mr._chopped, mr._silvair, rindo_chihaya, mizumiya_su, koganei_niko, kikirara_vivi, isaki_riona, kipfel_(vrchat), cinderella_(glass_princess)_(nikke), flutterpage, leviathan_(nikke), passionlip_(celeb_summer_experience_swimsuit)_(fate), luce_(jubilee_2025), sugar_lights_(umamusume), lappland_the_decadenza_(arknights), little_mermaid_(nikke), red_shoes_(nikke), dooby_(vtuber), lacrima_the_crimson_tears, peruere_(genshin_impact), yuuki_sakuna, la_manchaland's_don_quixote_(project_moon), yoshimi_(halloween)_(blue_archive), kazusa_(halloween)_(blue_archive), willow_(reverse:1999), mine_(idol)_(blue_archive), sakurako_(idol)_(blue_archive), katsuragi_izumi, dulcinea_(project_moon), sancho_(project_moon), ifa_(genshin_impact), iono_(fall_2024)_(pokemon), shirakami_fubuki_(saint), jimmy_(mouthwashing), kasane_teto_(halloween), fugue_(honkai:_star_rail), jia_xichun_(project_moon), daisuke_(mouthwashing), eupha_(metaphor:_refantazio), anya_(mouthwashing), fuyuan_(honkai:_star_rail), curly_(mouthwashing), louis_(metaphor:_refantazio), skully_j._graves, kiana_kaslana_(apho), kisaki_(kindergarten_uniform)_(blue_archive), eternal_sugar_cookie, kurumi_noah_(7th_costume), lucia:_pyroath_(pgr), quency_(escape_queen)_(nikke), rumi_(small)_(blue_archive), kashino_(fresh_and_sweet)_(azur_lane), sirius_(rabbit_of_paradise)_(azur_lane), bremerton_(pillowed_counseling)_(azur_lane), red_hood_(nonsense_red)_(nikke), gigi_murin_(1st_costume), elizabeth_rose_bloodflame_(1st_costume), cecilia_immergreen_(1st_costume), raora_panthera_(1st_costume), lumi_(wuthering_waves), moo_deng, anjo_nala, belfast_(blissful_service)_(azur_lane), amagi_(cv)_(azur_lane), cloud_strife_(bahamut_garb), tsuru_(love_live!), gonzales_(love_live!), lumine_(hot_springs)_(genshin_impact), c.i.e.l_(first_ascension)_(fate), yumia_liessfeldt, itokonoue_kaoru, ashiharatsumori_koito, ohitome_kohane, bb_cosmo_(fate), tepetlisaurus_(genshin_impact), yumkasaurus_(genshin_impact), irene_(the_adjudicatrix_and_her_resolve)_(arknights), tenochtitlan_(swimsuit_mooncancer)_(third_ascension)_(fate), tenochtitlan_(swimsuit_mooncancer)_(first_ascension)_(fate), rappa_(honkai:_star_rail), mizuo_haruichi, wu_song_(female)_(touqi_guaitan), naihe_(male)_(touqi_guaitan), naihe_(female)_(touqi_guaitan), brazilian_miku, theresa_apocalypse_(schicksal's_imperative), kaguya_(blue_archive), caesar_king_(zenless_zone_zero), lighter_(zenless_zone_zero), marina_(qipao)_(blue_archive), tomoe_(qipao)_(blue_archive), tenochtitlan_(swimsuit_mooncancer)_(fate), fukuyama_kenta, xu_fu_(swimsuit_avenger)_(fate), bb_dubai_(fate), dobrynya_nikitich_(swimsuit_lancer)_(fate), li_yuting_(mouse), burnice_white, zhu_yaoxing, lexington_(kancolle), marciana_(school_days)_(nikke), sophia_(manus_vindictae)_(reverse:1999), youhu_(wuthering_waves), space_ereshkigal_(first_ascension)_(fate), space_ereshkigal_(third_ascension)_(fate), space_ereshkigal_(fate), baobhan_sith_(exhibition_attire)_(fate), c.i.e.l_(fate), argus_(reverse:1999), tuesday_(reverse:1999), shi_zhaoying, papyrus_(arknights), narantuya_(arknights), pepe_(arknights), mogador_(kancolle), mini_durin_(genshin_impact), queen_dizzy, professor_niyaniya_(blue_archive), akemi_(blue_archive), sakura_miko_(flower_fantasista!), usada_pekora_(summer), hakui_koyori_(school_uniform), moze_(honkai:_star_rail), lingsha_(honkai:_star_rail), hakos_baelz_(steel_rose), gloire_(kancolle), feixiao_(honkai:_star_rail), asaba_harumasa, securityboo_(zenless_zone_zero), ajaw_(genshin_impact), ororon_(genshin_impact), kachina_(genshin_impact), chasca_(genshin_impact), mavuika_(genshin_impact), citlali_(genshin_impact), xilonen_(genshin_impact), jane_doe_(zenless_zone_zero), mualani_(genshin_impact), seth_lowell, mystic_flour_cookie, wanderer_(mega_mgc_coffee)_(genshin_impact), astellar_of_the_white_forest, luciana_de_montefio, qingyi_(zenless_zone_zero), tsukishiro_yanagi, asimov_(spindle_(aidenr0)), ottilie_(spindle_(aidenr0)), zhezhi_(wuthering_waves), nymph_(arknights), kinich_(genshin_impact), kirara_(phantom_in_boots)_(genshin_impact), nilou_(breeze_of_sabaa)_(genshin_impact), napoli_(azur_lane), admiral_nakhimov_(azur_lane), otomo_(cecilia_immergreen), shadow_milk_cookie, raziel_(summer)_(granblue_fantasy), abby_(wuthering_waves), rosanna_(chic_ocean)_(nikke), piper_wheel, fujii_ayame, uchida_momo, grem_(gigi_murin), bai_ling_(touqi_guaitan), sakura_(bloom_in_summer)_(nikke), mercuria_(reverse:1999), viper_(shine_of_love)_(nikke), konoka_(blue_archive), sparkle_(doll)_(honkai:_star_rail), calstone_light_o_(umamusume), durandal_(umamusume), rellana_twin_moon_knight, dokibird_(vtuber)_(bounty_hunter), kirino_(swimsuit)_(blue_archive), kanna_(swimsuit)_(blue_archive), fubuki_(swimsuit)_(blue_archive), chattino_(raora_panthera), aozora_kyoka, sylus_(love_and_deepspace), goldenglow_(party_in_the_garden)_(arknights), raora_panthera, cecilia_immergreen, gigi_murin, elizabeth_rose_bloodflame, tri_(zelda), you_jiujiu, nukumizu_kazuhiko, ziche_fuzhao, li_yuting_(female), li_yuting_(male), liu_qingcheng, mo_chengwei, fu_shun_(touqi_guaitan), huai_yunqing, firefly_(ar-26710)_(honkai:_star_rail), march_7th_(hunt)_(honkai:_star_rail), jiaoqiu_(honkai:_star_rail), yunli_(honkai:_star_rail), needle_knight_leda, emilie_(genshin_impact)
@Ikena Do you share a new Link?
@VKilko it would be again banned, its the same like i put here
Hi, is version 2.2 trained on the IL 0.1 or 2.0 model?
Compared to version 2.1fix, version 2.2 is more prone to errors in limbs, fingers, and toes under the same prompt
I compared 32 pictures and came to the same conclusion, the probability of finger deformity is much higher than 21fix.
After changing the prompt and continuing to test hundreds of pictures, I found that the difference was not particularly obvious in other prompts. It seemed that the difference was particularly serious only for the pictures that responded to spread pussy.
With the basic prompt set that I normally use, the 22 version produces pictures with almost no reasonable finger detail when using this prompt, while the 21fix version has very good finger detail.
Then I tested the 4x8 group of hand movements, such as v / waving, and their overall results were almost the same.
I think this may be because some new training set pictures below the horizontal line contaminated some prompts.
I pointed it out in "About this version" that Hands, including with it fingers, are worse, limbs on the other hand where the same on me. I don't tested toes much, not really liking feet. Also that fingers have less details, is because everyting looks a bit more anime like and less 2.5d like. So details like fingernails appearing less (but it really depents and fingernails is its own prompt, not using it can also remove them, since 2.2 is more prompt following).
Good
nice
Will you move the model, or maybe yourself, from Civitai? If so, where would you publish your work? Thanks.
Love it
I like the default style. Which style is it? Some other loras I have remove this model's default style, and I would like to stop that.
I love it
cool
Intricate details works too well, perfect
Great model, the only issue with it I think is it's quite iffy on making pit hair and things of that nature.
epic
pretty good
cool
👌
My favorite model, v2.2 is perfect, finally a waifu model which can deal with motion lines and motion blur without adding more Lora
Probably my new fave illustrious checkpoint, very accurate and consistent with character designs even with loras added, and has an easy time implementing two or more characters without shuffling traits between characters.
Very cool model!
Very often, when i try to generate a solo character, it generates a copy of it, fused together, even when using tags like "1girl, 1boy, solo" it still gives me a two headed mess 😭
Oh... the model works really poorly with 1516x2048 resolution.... a shame, I expected this from an XL model.
@Takujaba Model is trained on the normal sdxl resolution of bucket sizes of 1024. 1516x2048 is not the norm for a sdxl model. Even later Illustrious versions that was in parts trained with 1024 to 1536 can't handle 2048 resolutions that great.
yes,你找到解决办法了吗
QAQ
@Ikena perhaps, though waiNSFW works perfectly fine with this resolution all the time. Thanks for the heads up, i'll keep that in mind
I think in terms of creativity, and higher quality outputs v2.2 is pretty good, though it did get worse in terms of prompt adherence, with longer prompts, fairly creating a white woman for example when theres a dark skin tag, but none of those issues happen with v1.3 A, fingers also seem to have gotten a worse.
if you use "dark-skinned female" instead, is that problem still present? i actually prune "dark skin" if "dark-skinned female" is used in training. So maybe it wants to dark skin a male, that is not present
@Ikena ah dark skinned female did the trick, but I also noticed fingers being worse
@AndroidXL Hands/Fingers are worse, its really hard to get that right. I hope i get it better in the next update.
nice
can you tell us what changes there is within the versions? i just downloaded the Illustrious version yesterday and there's already a version updated TODAY lol. do i have to delete the old one or something like
Look at "About this version" on the right👍I don't updated anything in the last 14d
I do understand why 2.1 had to be abandoned, but it's still deeply unfortunate for me. It's so perfect for a lot of my use cases, and there are characters who I'm only satisfied with in that version.
totally agree with you
Pretty irksome to generate platinum hair without getting sharp teeth and pointed ears. Not everyone with light hair has to be a vampire.
what prompt do you use? there is no tag for platinum hair, "platinum blonde hair" exist and it don't create any vampire for me
if that possible, bro, can you somehow train model to understand more view angles, because, for example, "from side" is directly from side, because in english here is no precise words to say "third-person perspective", idk how to explain it actually, but camera angles are tough too work with
Model is trained with the images composition tags of danbooru https://danbooru.donmai.us/wiki_pages/tag_group%3Aimage_composition , some combinations works to give other angles
C'mon man, AI art is already as easy as it can get. Use an image describer such as Florence2 or Ollama Describer and find some art where the character is at the angle you want. It'll spit out the tags art like that is likely to have.
awesome details but the model from 2.2 made the head of every character super small.
erm wow, this is awesome
I say it with all my sincerity HASSAKU is the best model.but A little unsuccessful in multiple characters though
My favourite illustrious checkpoint, highly recommend
can u update characters data for v2? without changing style?
Tuyệt
Good model
the creativity and ddiverse outputs even with long prompts, is unmatched, its on a league of its own.
Pretty good
Am I the only one who has the problem that the female characters come out with red eyelashes or red eyes?
Nice
duplicate detection system of civit unpublished it, its now back. Also v2.1fix currently only creates noise on the generator
good work ,let the 2.1 version access to civitai generation, think is best one , all white picture with 2.1 fix !?? or my bad lora integration?
civit dev looks at it, not my fault
your work are great ,maybe conflictual illustrious /ponyXL lora ,no problem
@AIdundeeProd thanks, but its also not any conflict on your site. The generator simply put noise out currently. After it got republished
Model has gotten very inconsistent, similar to Mol Keun. Some results are perfect, most are turning into blank colors.
Stupid question, but is it somehow possible to make the model output more random results? Say if I don't specify location in the prompt, it will always generate some simple white studio background or some simple indoors room. Is there a way of making it so that model produces more "random" results, say, forest, or a castle, without specifying it in detail?
Or maybe there is some more "creative" model that I can use? I love the 1.3B
use wildcards to generate random background prompt tags
Version 2.2 makes the faces look too cute and anime-like, even when I prompt for realism. In contrast, version 2.1 fix gave me more realistic and well-shaped faces when I asked for realistic outputs
I hope future updates bring back that balance between realism and style.
Its intentional to be more rigged torwards anime, so it don't fall into realism on some prompts prune to it. While 2.1fix had that issue. So, 2.1fix is indeed better at realism prompted, because it fights less against its own more anime like style.
@Ikena Yeah, that’s exactly why I liked the 2.1 fix — it didn’t lean too much into realism or anime, it struck a really nice middle ground. That made it perfect for dark fantasy art and similar styles. I’ll definitely keep using it for those themes. Thanks for your efforts, though
Again, be aware: v2.1fix currently only creates noise on the generator since it got repuplished. Civitai knows it
Edit 30.05: It sometimes works, sometimes not
Check your positive prompts. Something or some styles could be doing that. Or some Lora.
2.2 seems to have worse background details quality, like it's producing watercolor-like background.
Great Model
On top
I think v2.2 is worse for generating hent4i than the previous version.......
I've got more deformed results in v2.2. compared to v2.1fix
For real it's awful, screws stuff too frequently, I thought it's my bad wording, but other illustrious-like models behave better
@GregovZuk WAI is still generally the best~
@Konstantin95 yep, it's just hard times
overall great model
Is this based on Illustrious 0.1 or 1.0?
0.1
@iloveoldmen oh
Despite being trained on 0.1, the result is phenomenally impressive. Even compared to WAI which is trained on 1.0, it's still generate superior image quality
Why does it sometimes generate a blank canvas? Is it because of tags or the model is bugged?
v2.1fix is bugged on the generator
this model has been disappeared at homepage, and cannot be searched out.
I have no problem to find my model? others also have no problem to find this model
hello,sorry to bother you,I really like the model you creat,it's so amazing!
but I can't generate image with this 2.1 fix version on civitai now,what happened?
For some reason it generates overly exaggerated forms: simple (big breasts) look like (gigantic breasts:2.7)
Sorry for the layman's question... I generally use wai, but what exactly are the differences between hassaku and wai... I tried it myself, but I couldn't quite figure it out...
Default style and the knowledge of more characters is the main difference
@Ikena
@Ikena I see...? Personally, I think hassaku is more flexible in expressing the prompts, although wai is also excellent. Maybe it's just my imagination...
I don't know... I guess you could say that “when a third party looks at the image, it feels consistent”...
Fantastic as it is! Waiting for new update though : )
And what do you want the new version to include that the current one does not?
I generated countless images with 1.3 Style A before just fine ... but lately somehow I generate deep fried noisy results, I really don't know why it happens but I believe I did something wrong. I'm struggling to find the fix to my problem.
Edit: Found the solution. I accidentally activated the Zero Terminal SNR button ... whoops.
Why are there jagged things around the lines of the pictures I generated using this model?
Both 2.1 and 2.2 have
Use "transparent background" "pixel art" and "jpeg artifact" in the negative. It can also be really sensitive to CFG if you use noise schedulers other than the one it was trained and tested on.
@some_cultist Thanks for the answer!
I used the magnification model to enlarge the first generated image and then sampled it again, and the abnormality disappeared.
75% of my generations with this model are monochromatic red. They have the stuff I ask for, but it's as though it's all through a red filter. I have tried putting monochrome red in the negative to no clear effect. Any idea what could be causing this? I'm running it in fooocus, as installed by Stability Matrix. I just downloaded the model, selected it, and started trying to generate. Anything else I should have done?
I had the same issue (but with yellow color) when adding badly trained loras. Does issue persist even without Loras?
might be your sampler/scheduler combo, or your cfg is too high/low. try adjusting the values in your sampler
good
through a lot of experimentation with many different models i have learned that for some reason hassaku is much better at producing duo images or multi-subject images in general. more specifically, a lot of other models will take prompt keywords and sort of randomly average them out on the subjects, while hassaku is much better at taking a certain keyword and only putting it on one subject in the image...sort of like hassaku "understands" there are multiple subjects but other models sort of think it's just one big extended subject
It'd be nice to have a style Lora version of this checkpoint
clean af
the best
this works so smooth!
Details
Files
hassakuXLIllustrious_v22.safetensors
Mirrors
hassakuXLIllustrious_v22.safetensors
hassakuXLIllustrious_v22.safetensors
hassakuXLIllustrious_v22.safetensors
hassakuXLIllustrious_v22.safetensors
hassakuXLIllustrious_v22.safetensors
hassakuXLIllustrious_v22.safetensors
hassakuXLIllustrious_v22.safetensors
hassakuXLIllustrious_v22.safetensors
hassakuXLIllustrious_v22.safetensors
hassakuXLIllustrious_v22.safetensors
hassakuXLIllustrious_v22.safetensors
hassakuXLIllustrious_v22.safetensors
hassakuXLIllustrious_v22.safetensors
hassakuXLIllustrious_v22.safetensors
hassakuXLIllustrious_v22.safetensors
Hassaku.safetensors
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.