Hello! Before continuing, I would like to warn you my descriptions are long. I'm not a fan of being cryptic about information. I don't want users to have to "connect the dots". I highly recommend that you take the time to read through everything. You'll learn the most that way. While I do enjoy kermit typewriting, I appreciate if people read everything before asking questions.<hr /><h3 id="this-lora-is-part-of-a-collection!">This LoRA is part of a Collection!</h3><hr /><h2 id="model-information">🔹 Model Information 🔹</h2>The following is a Character LoRA of Sango from Inuyasha: The Final Act. Here are the things that you should know. ◽️ This LoRA went through many revisions. I'm sorry if it annoyed anyone. ◽️ This LoRA has several versions, which are explained in the section labelled "Versions" ◽️ The dataset is mostly sourced from TFA. Yashahime does not contain good character designs. ◽️ While I want my LoRA to capture her face, hair and outfit as accurately as possible - I don't really care for anime accurate body proportions or 1:1 mirrors. Does that make sense? I don't want og retro anime but I also don't want the mature woman big booba style. I aim for the middle.<hr /><h2 id="versions">🔹 Versions 🔹</h2>I have multiple versions and I do believe I need to explain them in some depth.. SangoPonyV1 ◽️ I originally shared the wrong version of this in the past. ◽️ Trained with AdamW8bit with 1000 steps (100 images @ 5 epochs, 10 repeats, 4 batch size) ◽️ Accidentally trained with sdxl_vae ◽️ Dataset was trained with "anime coloring, retro artstyle" style tags. SangoPonyV2 ◽️ An almost identical mirror to V1. ◽️ Trained with Prodigy with 1500 steps (100 images @ 15 epochs, 4 repeats, 4 batch size) SangoPonyV3_a / SangoPonyV3_p ◽️ These two were a recent revision and use a different dataset, artificially increased and containing Miroku. ◽️ One version uses Adafactor, and the other uses Prodigy, which can be identified by the letter at the very end of the model name ◽️ Both were trained with 1230 steps (206 images @ 10 epochs, 2 repeats, 4 batch size) ◽️ Dataset was trained with "anime coloring, retro artstyle" style tags. SangoPonyV4_a / SangoPonyV4_p ◽️ This version is almost a carbon copy of V3, but with some differences. ◽️ The Adafactor version was trained with 1230 steps (206 images @ 10 epochs, 2 repeats, 4 batch size) ◽️ The Prodigy version was trained with 1030 steps (206 images @ 5 epochs, 4 repeats, 4 batch size) ◽️ I removed the "anime coloring, retro artstyle" style tags from the dataset. <hr /><h2 id="prompt-usage">🔹 Prompt Usage 🔹</h2>Here's everything worth of interest, or worth mentioning. There are no outfit triggers, such as "kimono_outfit" or "bodysuit_outfit". I haven't bothered with that yet. There is only the main trigger.. "sango". Appearance (Face) ◽️"sango, brown eyes, makeup, blunt bangs, brown hair, long hair" ◽️"sango, brown eyes, makeup, blunt bangs, brown hair, long hair, low-tied long hair" ◽️"sango, brown eyes, makeup, blunt bangs, brown hair, ponytail" ◾️ Optional tags: "black hair" (limited) Outfit (Kimono) ◽️"kimono, pink kimono, skirt, green skirt, fingerless gloves" ◾️ Optional tags: "sandals" (limited) Outfit (Bodysuit) ◽️"bodysuit, black bodysuit, shoulder pads, fingerless gloves" ◾️ Optional tags: "weapon, scabbard, mask, giant boomerang" (limited) All in One (Sample Prompts) ◽️"sango, brown eyes, makeup, blunt bangs, brown hair, low-tied long hair, kimono, pink kimono, skirt, green skirt, fingerless gloves" ◽️"sango, brown eyes, makeup, blunt bangs, brown hair, ponytail, bodysuit, black bodysuit, shoulder pads, fingerless gloves" I don't feel like mentioning the Miroku tags. Do note these are examples. Be sure to use score tags, expressions. <hr /><h2 id="tips-and-recommendations">🔹 Tips and Recommendations 🔹</h2>I'm not really a SD scientist, but I'm going to share a few points that may or may not help you. ◽️ Reference the sample images, be it the initial preview or the very NSFW extras I generate often. I do not scrub metadata, I don't do that shit. Everything is made with SD Forge ◽️ The sample images are a mix of same looking poses, and also varied x/y/z plots - it's for a reason. It's to see how well the model can perform. ◽️ All images use Prefect Pony v6.0. I don't care for SFW models, but I would say WAI and Hassaku are decent alternatives. ◽️ Use a weight between 0.8 and 1. ◽️ Use ADetailer. It doesn't work for me.

Sango (Inuyasha: The Final Act) | Pony XL | NSFW Support

I was running several tests in LoRA Easy Training Steps. I thought this was decent as a Prodigy test. I ran x/y/z plots and turned them into lazy gifs, they should give rudimentary guidance. See description. Network Rank/Alpha: 32/32 Optimizer: Prodigy, cosine Learning Rate: Unet/Tenc of 1 Others: Noise Offset of 0.05, Multires Noise Iterations of 6, Discount 0.3, SNR Gamma of 5 Dataset: 112 images, 5 epochs, 7 repeats, 4 batch size, 1000 steps + Keep Tokens, No Flip Augmentation, Shuffle Captions, No Regularization

Files with SHA256: e59025fd28446ae23f77842e5277618b164ac53883d2821eae5c07b985a81a9e