Sign up free, get 10,000 credits

Starter credits added when you sign up.

Notice: About Recent Slower Generation Speeds and Limited Computing Resources — Dear PixAI Community, Thank you for your continued love and support for PixAI. Recently, generation speeds on PixAI have…

View details

SDXL

PonyDiffusion_V6XL inpainting

PonyDiffusion_V6XL inpainting - AI Model cover image
Subido en
28 oct 2026, 14:55
Usos
74
Reseñas
Satisfactorio  (1)
Permisos
  • Permitir la generación y el intercambio de imágenes
  • Permitir que los usuarios descarguen este modelo
  • Usos comerciales

LoRAs recomendadas

Descripción

IMPORTANT INFORMATION I FORGOT TO ADDAMD Users have to have the arg --no-halfHere’s how you set the COMMANDLINE_ARGS environment variable across different operating systems.Windows (Command Prompt or Batch script):set COMMANDLINE_ARGS=--no-halfWindows (PowerShell):$env:COMMANDLINE_ARGS = "--no-half"Linux/macOS (Bash or Zsh):export COMMANDLINE_ARGS=--no-halfOther Unix-like OS (BSD, etc.):export COMMANDLINE_ARGS=--no-half(Less Common)Example for Fish shell:set -x COMMANDLINE_ARGS --no-halfFOR THE BEST RESULTS USE THE SAMPLER EULER AThe samples are from Forge/Auto1111 img2img though i did most of my testing using openOutpaint because i prefer it over the default plus outpainting is so much easier but i chose to upload the img2img samples because openOutpaint doesn't hold the metadata in the images you generate and just want to be as transparent as possible, whether you use img2img or openOutpaint or even comfyui (although not tested on,i don't see any reason it wouldn't work)..One thing i forgot to add for those that don't know, you should have your Inpainting conditioning mask strength set to 0.99 which you can do by going to settings and i believe it's in the img2img section, mines been moved with the Quicksettings list... So if you use openOutpaint like i do anytime it brings up that screen to change it to 1 just click no.This took me over 2 months of trial and error with so many merges and combinations trying to get a ponydiffusion inpainting model because as I'm sure for the ones that have tried to create a ponydiffusion inpainting model the normal way, it will generate images but its basically not usable. It was trained on sdxl so idk why it's the only model that does this and It's not just ponydiffusion but any fine-tuned ponydiffusion model has the same problem but i came across two fine-tuned pd models calledluminaverseandT-ponynai3 (V3.5)that was very very close to the default ponydiffusion model but the problem wasn't nearly as bad but still unusable as it was so these were the steps that i took if you want to for whatever reason recreate it yourself. you'll also need4thTailHentai_03which is another fine-tuned ponydiffusion model and lastly AnieaseXL_V2.0which is the only model that's not trained on ponydiffusion but i used it because its a pretty good model overall with characters it already knows but mostly for the way the inpainting model's color blends into your image its very clean and seamless This step idk if its mandatory but I'll list it, one test i did was i used the Model Toolkit Extension loaded the normal ponydiffusion_v6 model clicked on the advanced tab and then in the import i selected the t-ponynai3_v3.5 model and selected UNET-XL and clicked save and used this model in the merge i mention below (when i reference using the t-ponynai3 model, it was technically the ponydiffusion_v6XL model with a swapped UNET-XL)so using the checkpoint merger in Forge/Auto1111 i started by creating the first inpainting model setting Add difference with a strength of 1 checking Save as float16 this will prune the model and create an fp16 model (i tried baking in a vae and not baking one in and since i use a normal sdxl vae by default i didn't really notice a difference) for model A = sdxl.inpainting model B = luminaverse model C = sdxl base model (specifically sdxl.09vae)1. luminaverse.inpainting Now that you have that you'll need to create an aniease inpainting model using the same method above.2. Aniease.inpaintingthen you'll need the 4thTailHentail_03 model but this one doesn't need an inpainting model (i tried creating the inpaint model and the results were just as bad)3. 4thTailHentail_03so first i took the luminaverse.inpainting model and the aniease.inpainting model and merged them together using the checkpoint model merger Model A = luminaverse.inpaintingModel B = aniease.inpainting Weighted Sum of 0.3 which resulted in luminaverse-x-aniease.inpainting modelthen to add some of the pony back and also further help the color problem because the colors for the normal models are fine it's only after you create an inpainting model with them that it ends up the way that it does and i have no idea why... but i took luminaverse-x-aniease.inpainting and used checkpoint model merger again Model A = luminaverse-x-aniease.inpainting Model B = 4thTailHentai_03Weighted Sum of 0.2luminaverse-x-aniease-x-4thTailhentai_03.inpainting The results after this was alright, color was close to normal but it seemed to effect the inpainting model a little too much so i then used checkpoint model merger again Model A = luminaverse-x-aniease-x-4thTailhentai_03.inpaintingModel B = T-Ponynai3_V3.5.inpaintingWeight Sum = 0.5luminaverse-x-aniease-x-4thTailhentai_03-x-T-Ponynai3_V3.5.inpaintingso just a 50/50 split between the two which fixed the inpainting side of things but brought back the color issues so this is when i dropped checkpoint model merger and used the Model Mixer Extensionand im only listing the successful merge receipts if i listed every single merge i tried i would have to write a book on the process and no one wants that. and i would just like to stress im not an expert or even knowledgeable on Block Weights or the different things like Normal,Cosine,Inv.Cosine ect. all the tests that lead to this next merge were all trial and error and over 2 months of using different merging techniques and i almost gave up but thought i might try something i still don't fully understand, not even a little bit actually which is the DARE option that the extension model mixer recently added. and i also played around with the Adjust settings that it gives you which look like i also used the elemental merge which was also just trial and error im sure someone that actually knows about what these functions do can make a way better inpainting model, one were the artists styles are close to the same because they do still work but are just weaker as i'll show in my xyz plot.. that's just the default settings I'll just post the settings that were in the Metadata but if you just wanted to copy the settings over this part specifically is the "ModelMixer adjust" part which i bolded the important parts but what worked was Model A = luminaverse-x-aniease-x-4thTailhentai_03-x-T-Ponynai3_V3.5.inpainting (i had a different name for it but it's the same as the Final-TestPonydiffusion_V6XLMix.inpainting model)Model B = 4thTailHentaiModel_03.safetensorsModelMixer merge mode b: DAREModelMixer alpha b: 0.5ModelMixer mbw mode b: TrueModelMixer use elemental b: TrueModelMixer calcmode b: Simple CosineModelMixer mbw b: ALLModelMixer mbw weights b: "0,0.125,0.15,0.125,0.15,0.2,0.25,0.2,0.15,0.2,0.25,0.125,0.15,0.125,0.15,0.2,0.25,0.2,0.15,0.2"ModelMixer elemental b: "IN00-IN03 OUT06-OUT08:skip_connection out_layers op in_layers emb_layers:0.25"ModelMixer model a: Final-TestPonydiffusion_V6XLMix.inpainting.safetensors [9fefd616ba], ModelMixer max models: 3, ModelMixer adjust: "0.05,0.05,0.15,0.1,0.025,-0.01,0.025,-0.15", ModelMixer use model b: True, ModelMixer model b: 4thTailHentaiModel_03.safetensors [a1f96244d3], ModelMixer merge mode b: DARE, ModelMixer alpha b: 0.5, ModelMixer mbw mode b: True, ModelMixer use elemental b: True, ModelMixer calcmode b: Simple Cosine, ModelMixer mbw b: ALL, ModelMixer mbw weights b: "0,0.125,0.15,0.125,0.15,0.2,0.25,0.2,0.15,0.2,0.25,0.125,0.15,0.125,0.15,0.2,0.25,0.2,0.15,0.2", ModelMixer elemental b: "IN00-IN03 OUT06-OUT08:skip_connection out_layers op in_layers emb_layers:0.25", ModelMixer use model c: False, ModelMixer use model d: FalseAfter this merge i ended up with Mix100+Test4FixPonydiffusionXL.inpainting.safetensorsBut still not done yet, i had to be sure it was basically still the pony model which has all the artist styles (over 750+ I'll post below) and we all know pony is a perfect nsfw model so that was the first test i did using different artists tags and I'm actually still shocked that for the most part work. I can't say they all work as i couldn't test all of them it would have taken to long, all the ones i did test worked so i figured It's good enough.Last step - used Supermerger extension to merge 4thTailHentai's Lora Extract with the inpainting model which seemed to really help bring out some of the artist styles that were really hard to tell if they had any effect or notTHIS MODEL IS NOT USED FOR GENERATING IMAGES, FOR THAT USE THE NORMAL PONYDIFFUSION_V6XL MODELSorry I forgot the artist list and i tried to upload the txt file but don't even know if you can on civitai so sorry in advance for having to scroll so muchSteps to Format this list is after it at the bottom. It's only 2 steps using Notepad++ Which I'll Screenshot.matsunaga_kouyou m-da_s-tarou cutesexyrobutts gweda fumihiko_(fu_mihi_ko) kantoku cle_masahiro kame_(kamepan44231) homare_(fool's_art) 92m personal_ami sciamano240 kuro_kosyou bigrbear samip wisespeak asutora hayabusa fujima_takuya varniskarnis 40hara seraziel kairunoburogu nottytiffy asanagi vahn_yourdoom kase_daiki cluseller iwbitu themaestronoob syhan niliu_chahui ibuki_notsu tama_(tamago) ke-ta scottie_(phantom2) ricegnat ishikei as109 abmayo lvl_(sentrythe2310) mignon chita_(ketchup) nekoda_(maoda) aki99 tab_head mashuu_(neko_no_oyashiro) ndgd doitsuken6 chahan_(fried_rice0614) kukie-nyan beijuu yoako harris_hero missile228 nagishiro_mito mao_lian_(nekokao) puuzaki_puuna yoshizawa_tsubaki yukito_(dreamrider) hu_dako omone_hokoma_agm kfr alp sabamen deaver ouma_tokiichi ruu_(tksymkw) bluefield sukemyon tachimi_(basue) kisaragi_nana sakuraba_hikaru_(loveindog) casino_(casinoep) unfairr null_(nyanpyoun) mochirong koyuki_(kotatsu358) sasaame pottsness miyase_mahiro ergot cowfee re:shimashima shpo tony_guisado roropull everyntrge yuano hater_(hatater) monobe_yuri takunomi z.taiga jeneral camonome nekoya_(liu) sayori_(neko_works) aoi_ogata kaedeko_(kaedelic) himitsu_(hi_mi_tsu_2) misaka_12003-gou neko_neko_koneko ratatatat74 afrobull rokugou_daisuke masami_chie huke ushiyama_ame mathias_leth sakura_chiyo_(konachi000) cirilla_lin xiujia_yihuizi umakuchi_shouyu deel_(rkeg) masabodo chanta_(ayatakaoisii) sendo_(sendrawz) saburou_(hgmg) tateha_(marvelous_grace) nanoless sub-res wlop sogaya ge-b ateoyh idaten93 dearmybro

Comentarios

Obra de arte hecha por PonyDiffusion_V6XL inpainting