BancinXL

V3! As I am working on a new Pony model for fun, I did a quick side-quest and created a V3 for BancinXL! Based on Juice and V2 with some LoRA and MeichiDarkmix checkpoints, i'll do a longer story later 😉 (PS: sorry for the messy showcase pictures, it's been a while since i generated Pony stuff and i forgot how much it is difficult to get good hands with it) TL;DR - is the main Pony version, including the "classics" LoRA (InCase/Vixon/Expressiveh/Disney) - is variant of V2, more shiny, more crazy Juice! is like a hard liquor, tasty but drink it pure and you'll go blind ;-) I have told in many posts that i love these models, but sometimes, working with them is not yielding what i wanted (but maybe, what i deserved :D). The Juice version of BancinXL is simply a 50/50 merge mix of PsychoNex and V2. Not intended to be the main version, but still a whole lot of fun for more chaos :D Go check work! V2 Rework of the V1, with a similar approach, with just a few more LoRA to give it some oomph. Still based on , before jumping on the Illustrious/NAI XL trend :D Recipe: BancinBase (using SuperMerger): - (1.0) [Tried the Pony version, V1 still the better fit] - (0.65) ❤ - (0.55) ↑ 0.05 - (0.35) ↓ 0.1 - (0.25) 🆕 [After dimension rework from 16 to 32] - (0.2) 🆕 - (0.15) 🆕 - (1.2) 🆕 [After dimension rework from 16 to 32] BancinTemp (using Checkpoint Merger): - 0.8 x BancinBase + 0.2 x BancinXL V1 - Baked in VAE: (switch from SharpSpectrum, still good colors, just less strong) BancinXL_V2 (using SuperMerger): - BancinTemp + (0.4) x Usage: I have added a step/sampler comparison to help choose, but the model works good with the classics without any issues. Euler a give more smooth picture while DPM++ 2M give more crisp results. For Euler a, 40 steps give a better looking result, while only 20 steps are enough for DPM++ 2M. DPM++ 2M SDE change the composition a bit but is nice for experiments. I tend to use CFG between 4 and 6 to give more or less liberty to the generation, but that's a personal preference. Regarding prompts, most are done with the classic Pony scores stuff, here is what i do for reference: Expressiveh, (optional LoRA trigger words), masterpiece, best quality, PonyScores7, , (character name as recognized by Pony), (body description), (scene description), (background and lighting) For negative prompt, except for specific, it's always this: PonyScoresNeg, (worst quality, low quality:1.4), (deformed, distorted, disfigured:1.3), poorly drawn, bad anatomy, wrong anatomy, extra limb, missing limb, floating limbs, bad hands, missing fingers, extra fingers, fused characters, muscular female, deep_negative_pony Embeddings used (real SDXL TI, SD1.5 TI are useless unless you hack stuff in ComfyUI, yes, i am looking at you EasyNegative ^^;): - - - Personal trick: I personally prefer the crispiness of DPM++ 2M for the overall picture, but faces are generally better with Euler a. To mix the best of both, I generate using DPM++ 2M with 40 steps and when kicks in, it is configured to switch only the sampler for the face to Euler a (will keeping steps, prompts, clip skip and so on). Simple and does the trick, didn't want to bother with also choosing CFG/Steps/etc...in the extension. :D V1 So. A very simple merge to integrate the LoRAs the amazing creator use in almost all his/her pictures ^^; (GO CHECK HIS/HER WORKS!) Made to allow trying his "style" on my GPU poor setup without overloading it too much. Base model: LoRAs and weights: - (1.0) - (0.65) - (0.45) - (0.5) Update: Wow, scrolled the Disney Princesses Gallery and discovered everyone and more are using approximately the same mix of lora ha ha ha. Lots of VRAM consumed on merging those with Pony each time instead of using a pre-baked one ^^;