AnBan_Illus

Shin V1Rework of AnBan from Quirk V3 using Alter V3 and HoJ V2 together with a bit of Ri-mix to get back character semi-realistic style, lost to a more "flat" anime like style during Illustrious 2.0 upgrade.Quirk V3Quick update based on Illustrious 2.0. Not too complicated merge this time (well still, a bit), main goal was to keep the style but get the composition and CLIP "almost" Illustrious vanilla.Vpred V2So, not what you were expecting as my next release, isn't it? :DJust a quick fix of the V1 using HoJ dataset to boost it a bit as an experiment.Provided as-is (some issues with hands are back T_T)Quirk V2Fixing QuirkV1, long storyQuirk V1Yo Dawg, i heard you like Illustrious so we have put some more Illustrious in your Illustrious checkpoint 😂Quirk V1 is a quirky merge of some Illustrious V1.0 in Alter V3 (with a touch of AnBan 1.0 for stabilisation) for the sake of tests and fun. Nothing serious here, but it gives a bit more details and shadows.Alter V3Probably the last iteration of this model before moving on in an other direction, lots of work went in fixing issues the previous version had. Thanks to everyone for the feedbacks!The long story is here and here. Regarding usage, make sure to check this image to find the couple sampler/scheduler which fit your fancy :DAlter V2With all the kind words and great stuff made using AnBan, i decided to build a V2 of AlterBan, and to go in the opposite direction of BancinIXL: smooth and calm rather than vivid and wild :DThis version count also as an experiment to test my toy LoRA "Ponyfier". To build Alter V2, i leveraged a gold old Pony classic: the 5.2 version of DucHaiten Pony.Using Supermerger, i did a "Duc minus Pony" LoRA that i merged together with my Ponyfier to create a kind of Illustrious style LoRA that i then tested and merged with AlterBan.The Alter V2 version is giving good results (even if hands are still a bit tricky, i'll need to build for myself an all-purpose hand fixer LoRA), still get Illustrious LoRA working quite nicely, and i leveraged a trick that was gifted to me: don't use quality tags in the positive prompt. This help getting 1boy working better and overall don't mess the generation :DI am using this setup for the showcase:Euler a + Automatic40 stepsCFG 6Adetailer 40 steps too, denoising strentgh 0.2, other settings at default, VAE: NoneThanks for reading!NB: since i don't use "artists" style tags, i don't test those and i am pretty sure my models are messing this kind of stuff, so, not the best if you are relying on anything like this.Older version history below:Update: important information regarding AlterBan (and possibly all my models using custom VAE)I have noticed that using the dark VAE mess up generation when using Adetailer (even more than with other VAE). It's probably due to the fact that the VAE re-apply a "color/contrast" modifier on a already modified picture, leading to this infamous "box effect".If this happens too much, switch back the VAE to standard SDXL VAE either from the start (in A1111/Forge, edit the checkpoint parameter and choose a preferred VAE) or during Adetailer (option use a separate VAE).Alter!I am ALWAYS open to feedbacks and criticism and all inputs are welcome because they can only help me cover areas where i am less confortable to get better ❤That's why, here is a alternate version of AnBan, based on V1 and finetuned with a dataset built from the Vpred version with contrast and color dimmed for a softer look, more adapted to some more mysterious or cozy pictures. While doing this, i messed up the hands and fixing those brought back some more contrast, but i'll probably do a separate LoRA (something like "AnBan_Noir") for a more Gothic/Film noir ambiance :DIn this version, i switched the baked in VAE from flatpiece to dark which also help mute colors and all showcase pictures have been done using Euler a instead of DPM++ 2M with "vivid colors" in the negative prompt. (note that the VAE will kick off at the end of the generation, preview pictures during generation will be brighter that the end result).Vpred V1!The TL;DR and long story are here: https://civitai.com/articles/10620V2AnBan_Illus V2 is here! Probably the last version of this model before building a fully Illustrious BancinXL :D (NB: AnBan is an Epsilon Prediction model).All infos on the making here: https://civitai.com/articles/10148/bancinxl-for-illustrious-step-1V1Introducing Ante Bancin illustrious, aka AnBan_Illus! I honestly can't say "yeah, everybody should use this!" as this checkpoint is more of a "personal" resource for my own kinks and to serve as a base for training the "BancinXL" style upon :DI tested and used with success for now:DPM++ 2M - KarrasClassic SDXL resolutions: 832x1216, 1024x1024, 1216x83225 steps, CFG at 5Adetailer (only if needed) at 0.4 for inpainting, sometimes switching to Euler a if i want a softer feel.Hires.Fix at 0.4 too, with 20 steps, CFG at 5 tooForge integrated FreeU and SelfAttentionGuidance with default valuesand i kept my CLIP skip at 2, probably because i am lazy :DPositive prompt:[whatever your usual prompt is, without the Pony Scores], masterpiece, best quality, amazing quality, very aesthetic, absurdres, newestNegative prompt (if you want to remove the flat touch, use flat color in the negative prompt and bonus point, deep_negative_pony embedding give good result too!)bad quality,worst quality,worst detail,sketch,censored,watermark,signature,artist name,flat color,deep_negative_ponyOrigin storyYou can probably already close the tab and go back to your own activities, but if you like stories, here is one ;-)Everyone is jumping on the Illustrious train and freckledvixon is spamming with cool LoRAs i can't test until i jump too. :DBut for now, everybody and everyone is releasing their own trained/merged checkpoints everywhere (even if some of those are starting to get lots of attention) and there was not one checkpoint that exactly ticked with what i like... so, like all the people, i did my magic mixture to get a base i like for tests and on which i'll build later on to port BancinXL to Illustrious.Why this mixture? Here is some example:IllustreijL (i like the body type, more mature, but the facial expression is not "anime like" and the skin texture is not "shiny smooth", even with a LoRA)NTR-MIX (i like the facial expression, more anime like character, but not the body type and too "flat")Lilly Anime (much more close to my taste on textures, but the facial expression and body type is not quite what i wanted)So, what could i do? Guess-mix stuff! Here is the recipe with the resources from above (and this LoRA), because in the full spirit of Opensource, i don't believe in hiding the original sources, not crediting the authors and claiming this is "mine" :DCheckpoints are merged using Checkpoint MergerLoRA are merged in checkpoints using SuperMergerAnd here is what i ended up with :DStill quite not there, too light for my taste, but it will be perfect for my next steps ;-)Thanks for reading!