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

One of the more recent versions. Trained on site. I forgot what _m stood for. This dataset includes Miroku in them. It lacks style tags, which I think is more proper. This one contain backgrounds that are like mostly anime and then some simple backgrounds. Network Rank/Alpha: 32/16 Optimizer: Adafactor, cosine with restarts, 3 restarts Learning Rate: Unet 0.0005, Tenc 0.00005 Others: Noise Offset of 0.1, Multires Noise Iterations of 6, Discount 0.3, SNR Gamma of 5 Image Info: 120 images, 4 epochs, 9 repeats, 4 batch size, 1080 steps + Keep Tokens, Flip Augmentation, Shuffle Captions, No Regularization

SangoPonyXL_m.safetensors

120 images (without styles).zip

SangoPonyXL_m

SangoPonyV4_p

SangoPonyV4_a

SangoPonyV3_p

SangoPonyV3_a

SangoPonyXL_styles

SangoPonyV2

SangoPonyXL_prodigy_e5

SangoPonyV1

If you like my work and would like to leave a review, I mean.. feel free to. I do this because I'm a degenerate but also do this for fun.<hr /><h3 id="this-lora-is-part-of-a-collection!">This LoRA is part of a Collection!</h3>Sango (Pony XL) Sango (Illustrious XL) -- you are here Sango (Pony XL/Illustrious XL + Doujins) soon<hr /><h3 id="lora-information">Lora Information</h3>Sango from Inuyasha: The Final Act. Dataset is mostly from that anime. I've re-used datasets and made an attempt to make Illustrious versions. Two versions right now, likely final. They're actually not too different. SangoIllustriousXL_m - Uses a dataset of 120 images - Slightly unrefined, contains tags like "black hair, brown hair" for most. SangoIllustriousXL_m2 - Uses a dataset of 206 images - Slightly refined. Fixed some tags. - More Epochs, same steps. TE LR of 0.0001 I trained a Prodigy version but to be honest I don't care right now. I'm not a current fan of Illustrious XL - personally out of inexperience. PSA. This is not going to mirror the anime, or have anime proportions. Don't expect it. <h3 id="prompt-usage">Prompt Usage</h3> Trigger Word: "Sango" Face/Appearance: "brown eyes, makeup, bangs, black hair or brown hair" Hairstyle: "long hair, low-tied long hair" or "long hair" or "ponytail" Kimono: "kimono, pink kimono, skirt, green skirt" + "fingerless gloves, sandals" (optional) Bodysuit: "bodysuit, black bodysuit, shoulder pads" + "fingerless gloves" ? (optional) Extras: "sword, scabbard" or "weapon, scabbard" for the sheathed katana / "giant boomerang" for the hiraikotsu / not well tagged or understood at the moment. Sample Prompts<pre><code>Sango, 1girl, solo, smile, brown eyes, makeup, bangs, brown hair, long hair, low-tied long hair, breasts, kimono, pink kimono, skirt, green skirt, fingerless gloves, looking at viewer, waving at viewer, standing, cowboy shot, outdoors, day, village</code></pre><pre><code>Sango, 1girl, solo, smile, brown eyes, makeup, bangs, brown hair, ponytail, breasts, bodysuit, black bodysuit, shoulder pads, looking at viewer, hand on own chest, standing, cowboy shot, outdoors, day, village</code></pre><pre><code>Sango, 1girl, solo, annoyed, brown eyes, makeup, bangs, brown hair, long hair, breasts, bodysuit, black bodysuit, shoulder pads, looking at viewer, crossed arms, standing, cowboy shot, outdoors, day, village</code></pre>Do note these are examples. Be sure to use quality tags, expressions.. probably make use of BREAK if your prompts are over 75 tokens. I'm currently learning about that.<h3 id="tips-recommendations">Tips / Recommendations</h3>I recommend a weight of 1, CFG above 5. I am unsure about 1024x1024 vs upscaling. Use Adetailer. You can probably reference the gallery. At the time of writing it is 20 images strong. Models? I've used three in the gallery, they are <a target="_blank" rel="ugc" href="https://civarchive.com/models/827184?modelVersionId=1761560">WAI-NSFW-Illustrious-SDXL</a>, <a target="_blank" rel="ugc" href="https://civarchive.com/models/992378?modelVersionId=1111838">Prefectious XL NSFW</a>, and <a target="_blank" rel="ugc" href="https://civarchive.com/models/1224788?modelVersionId=1873831">Prefectious Illustrious XL</a>. Last two are similar. Prefect's style is to "booba-ify" characters. Thick proportions. I have no issue with it personally. WAI is a good solid pick. I haven't tried Hassaku's IL variant, could try that.<hr /><h3 id="enjoy!">Enjoy!</h3>

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

Trained using Civitai (500 buzz), after using Lora Easy Training Scripts for reference. Settings: Network Rank/Alpha: 32/16 Optimizer: Adafactor, cosine with restarts, 3 restarts Learning Rate: Unet 0.0005, Tenc 0.00005 Others: Noise Offset of 0.1, Multires Noise Iterations of 6, Discount 0.3, SNR Gamma of 5 Dataset: 120 images, 4 epochs, 9 repeats, 4 batch size, 1080 steps + Keep Tokens, Flip Augmentation, Shuffle Captions, No Regularization

SangoIllustriousXL_m.safetensors

Sango (Inuyasha: The Final Act) | Illustrious XL | ???? Support

SangoIllustriousXL_m

SangoIllustriousXL_m2

Files with SHA256: d04c3610ac06d9b29a3bb1e7d752b1168508313e882f7da7f75a24117d2a8843