<h1 id="announcement:-xm4uuq9ps">Announcement:</h1>It's been a very long time since I last updated this model. It's my "first-born," and I'm very proud of it. However, unfortunately, I no longer have the capacity to maintain this model. If you like it and want to develop it further, feel free to fork the model and customize it as you see fit—I would be very happy. Please don't forget about me. See ya! (●'◡'●)Don't forget me.I have a secret project, you can follow my Instagram:<a target="_blank" rel="ugc" href="https://www.instagram.com/6789zkm/">BlueCat (@6789zkm) • Instagram photos and videos</a><h2 id="introduce:-tj5de6rsc">Introduce:</h2>I don't know how to introduce it, but it's been renamed several times. Project-K is an open, free to use and fine-tune AI-art model. It was created by my curiosity. Hope you will like it. Have fun! (●'◡'●).<h2 id="tips:-fuia6rzmi">Tips:</h2><ul><li>The `masterpiece` and `best quality` tags are not necessary, as it sometimes leads to contradictory results, but if it is distorted or discolored, add them now.</li><li>The CGF scale should be 7.5 and the step counts 28 for the best quality and best performance.</li><li>Use a sample photo for your idea. Interrogate DeepBooru and change the prompts to suit what you want.</li><li>You should use it as a supportive tool for creating works of art, and not rely on it completely.</li><li>Normal: Clip skip = 2.</li></ul><h2 id="desires:-h4gl2i5jq">Desires:</h2>As it is a version made only by myself and my small associates, the model will not be perfect and may differ from what people expect. Any contributions from everyone will be respected.Want to support me? Thank you, please help me make it better. ❤️<h2 id="faqs-that-i-made-up-(might-be-helpful-for-you).-zr5tvh2t1">FAQs that I made up (might be helpful for you).</h2>Do you have a Discord server?<ul><li>Well... No, I'm not prepared for that. Besides, my English is not very good.</li></ul>Can I use your model to make money?<ul><li>Of course, feel free to do so. But you have to promise me not to sell my original model. If I'm mentioned in your credits, I'll be happy.</li></ul>Why is your model's size 5.19 GB instead of 7.14 GB? I have OCD. Ahhhhhhh<ul><li>Well... it relates to the EMA training process. In order to maximize cost efficiency, we needed to train a model using fp16. However, when I trained the model using d8ahazard's dreambooth, enabling fp16 with EMA resulted in such a skewed model. Boom, we have fp16 EMA and fp32 UNet.</li></ul>Can you help me improve the model's output quality?<ul><li>Based on my experience, you should use Easy Negative, set Clip skip to 2, and make the Negative prompt similar to the images in the preview. You can also try using the hires fix Lantent once; it works fine on my potato computer. Lastly, use Adobe Photoshop and a bit of luck.</li></ul>Can you help me add the things I want to your model?<ul><li>Of course, please comment below. I will do it for you for free to the best of my abilities.</li></ul>Is there a secret?<ul><li>Secrets are not revealed, but you may have already figured it out. Bro, we're quite similar =))).</li></ul>Of course, there are more questions. Comment below, and I'll answer you or someone else will. Good luck.<h2 id="special-thanks:-this-wouldn't-have-happened-if-they-hadn't-made-a-breakthrough.-0663q1k7z">Special Thanks: This wouldn't have happened if they hadn't made a breakthrough.</h2><ul><li><a target="_blank" rel="ugc" href="https://huggingface.co/runwayml/">Runwayml</a>: Base model.</li><li><a target="_blank" rel="ugc" href="https://github.com/CompVis/">CompVis:</a> VAE Trainer.</li><li>stabilityai:<a target="_blank" rel="ugc" href="https://huggingface.co/stabilityai/sd-vae-ft-mse-original"> stabilityai/sd-vae-ft-mse-original · Hugging Face</a></li><li><a target="_blank" rel="ugc" href="https://github.com/d8ahazard/sd_dreambooth_extension">d8ahazard</a>: Dreambooth.</li><li><a target="_blank" rel="ugc" href="https://github.com/AUTOMATIC1111/">Automatic1111</a>: Web UI.</li><li><a target="_blank" rel="ugc" href="https://github.com/Mikubill/">Mikubill</a>: Where my ideas started.</li><li>Chat-GPT: Help me do crazy things I never thought I would.</li><li>Novel AI, Anything Model, Abyss Orange Model, CentusMix Model, Grapefruit Model: Dataset images. An AI made me thousands of pictures without worrying about copyright or dispute.</li><li>Danbooru: Help me write the correct tag.</li><li>My friend and others: Get quality images.</li><li>And You 🫵❤️</li></ul><h1 id="copyright:-iq9c2v4w5">Copyright:</h1>This license allows anyone to copy, and modify the model, but please follow the terms of the CreativeML Open RAIL-M. You can learn more about the CreativeML Open RAIL-M <a target="_blank" rel="ugc" href="https://huggingface.co/spaces/CompVis/stable-diffusion-license">here</a>.If any part of the model does not comply with the terms of the CreativeML Open RAIL-M, the copyright and other rights of the model will still be valid.All AI-generated images are yours, you can do whatever you want, but please obey the laws of your country. We will not be responsible for any problems you cause.Don't forget me.<h1 id="have-fun-with-your-waifu!-('')-gh6mnuog6">Have fun with your waifu! (●'◡'●)</h1>I'm not sure if you've read this section, but if you're wondering, "Where are versions 1.0 and 2.0? What happened to them?" Well, don't fret, they've been retired because they were about as effective as a chocolate teapot. I wouldn't want you to waste your time downloading something that's about as useful as a screen door on a submarine.This model has a bit of a past... it's kind of hard to describe, it's like a mix of comedy and melancholy. I usually call it the "awkward situations" mode. It was trained on an utterly illogical system, an RTX 3050, yes, you heard that right, an RTX 3050. Its speed is, well, let's just say it makes snails look fast. Training for one epoch takes a whopping 4 hours, OMG. And the results? Utterly disastrous. Sometimes, I question why I'm doing this when I could be downloading something better made by someone else. Yet here I am, still at it, up until now.Simply put, it's all for the thrill, the sensation of creating something, even if it's hilariously terrible. It's like making a lopsided cake and still being proud of it. So, if you're like me, here's a piece of advice: "Where there's a will, there's a way," so keep on truckin' in your career.

Project-K (Kawai)

Feature:<ul><li>Add Guard Info to the model.</li><li>Ema-only</li></ul><ul><li>Fix color loss.</li><li>Image quality is greatly enhanced. Thank you, my friend.</li><li>Hentai improved—still my friend.</li><li>Kawai Diffusion's most powerful ability is "enhance" (img2img). It will make a bad photo look better.</li><li>True "kawaii desu"... Haizzzzzzz</li></ul>Limit:<ul><li>The drawing is hard, not soft.</li><li>The hand is still bad.</li></ul>Training:<ul><li>Data: Created by another AI.</li><li>Schedule: DDIM.</li><li>Optimizer: AdamW.</li><li>Precision: FP32.</li><li>Hardware: Google Colaboratory Pro - NVIDIA A100 40GB VRAM, TESLA V100-SXM2 16GB.</li></ul>Guard:<ul><li>{'name': 'Kawai Diffusion', 'author': 'Huy Nguyen Quang - Ojimi - Kawai Project', 'security_key_md5': '45de2f255cb99000be4d4b769255b6f0'}</li></ul>

projectKKawai_v30LTSEmaonlyOUTDATE.yaml

projectKKawai_v30LTSEmaonlyOUTDATE.safetensors

v3.0-LTS-emaonly[OUTDATE]

5.1.0-Fabric

5.0.1-250723

pre_nightly130523

pre_sparkle100523

v4-charm_LTS

sparkle453[OUTDATE]

v3.2[OUTDATE]

v3.1[OUTDATE]

v3.0-LTS-pruned[OUTDATE]

Feature:<ul><li>Add Guard Info to the model.</li></ul><ul><li>Fix color loss.</li><li>Image quality is greatly enhanced. Thank you, my friend.</li><li>Hentai improved—still my friend.</li><li>Kawai Diffusion's most powerful ability is "enhance" (img2img). It will make a bad photo look better.</li><li>True "kawaii desu"... Haizzzzzzz</li></ul>Limit:<ul><li>The drawing is hard, not soft.</li><li>The hand is still bad.</li></ul>Training:<ul><li>Data: Created by another AI.</li><li>Schedule: DDIM.</li><li>Optimizer: AdamW.</li><li>Precision: FP32.</li><li>Hardware: Google Colaboratory Pro - NVIDIA A100 40GB VRAM, TESLA V100-SXM2 16GB.</li></ul>Guard:<ul><li>{'name': 'Kawai Diffusion', 'author': 'Huy Nguyen Quang - Ojimi - Kawai Project', 'security_key_md5': '45de2f255cb99000be4d4b769255b6f0'}</li></ul>

projectKKawai_v30LTSPrunedOUTDATE.yaml

projectKKawai_v30LTSPrunedOUTDATE.safetensors

v4-charm LTS version (snapshot 16-04-2023).Thank you for supporting us all this time.Feature:<ul><li>Full of Kawai Diffusion v4.</li><li>Support long-term use.</li><li>High-end quality.</li><li>Good compatibility with most Lora, Textual Inversion.</li><li>Model with mixed precision:<ul><li>Unet: fp32 (3.6 GB)</li><li>Ema: bf16 (1.6 GB)</li></ul></li></ul>Limit:<ul><li>Error, bad, bad hands, color loss.</li><li>The compatibility with some Lora, Textual Inversion... is considered to be ineffective.</li></ul>⚠️ Usage warning:<ul><li>The output images may not be suitable for all ages, please use with additional safety measures. (violence, sexual)</li><li>Some parts of the model may not be suitable for you. Please use it with caution.</li></ul>

projectKKawai_v4CharmLTS.yaml

projectKKawai_v4CharmLTS.safetensors

Let me know how you feel and rate it in the comments section. And what do you expect for the next version?Feature:<ul><li>Recommended for inpainting, image to image.</li></ul><ul><li>hard, nonguard is a special tag of Kawai (very different from nsfw), extremely powerful.</li><li>More content.</li></ul>Limit:<ul><li>Not the LTS version. May change in the future.</li><li>The drawing is hard, not soft.</li></ul>Training:<ul><li>Data: Created by another AI.</li><li>Schedule: DDIM.</li><li>Optimizer: AdamW.</li><li>Precision: FP32.</li><li>Hardware: Google Colab Pro - NVIDIA A100 40GB VRAM.</li></ul>Guard: None

projectKKawai_v32OUTDATE.safetensors

projectKKawai_v32OUTDATE.yaml

<h1 id="the-fallen-angel-anime-character-generation-via-rise-scheduler-1zsk0sq61">🧪 The Fallen Angel — Anime Character Generation via RISE Scheduler</h1><h2 id="abstract-lqzziogba">Abstract</h2>The Fallen Angel is a custom-trained anime checkpoint designed for simplified prompt input while maintaining high-fidelity, lovable anime girl characters. The model introduces a novel training scheduler—RISE (Rise Inverse Stable Evolution)—as part of ongoing research into scheduler behavior within DreamBooth. This is the 89th training iteration in my series, and the resulting model performs well with simple poses, high-resolution upscaling, and minimal prompt engineering.<h2 id="introduction-6wnghkqdr">Introduction</h2>Anime generation in diffusion models often requires complex prompt crafting and intensive tuning. The Fallen Angel seeks to simplify this process by introducing a bias towards well-composed, single-character (1girl) outputs with minimal prompts. By emphasizing simplicity and character fidelity, this model is particularly suited for enthusiasts and creators who prefer consistent, charming results without heavy prompt engineering.<h2 id="methodology-kekbjy6ur">Methodology</h2><h3 id="model-focus:-8klp5ghlm">Model Focus:</h3><ul><li>Style: Anime (1girl, clean pose, expressive characters)</li><li>Prompt Efficiency: Reduced prompt complexity</li><li>Orientation: Primarily 512x768 (portrait)</li></ul><h3 id="training-details:-2o1tt5o8s">Training Details:</h3><ul><li>Training Iteration: #89</li><li>Custom Scheduler: RISE (Rise Inverse Stable Evolution) <a target="_blank" rel="ugc" href="https://www.hitoha.moe/rise-algorithm-for-dreambooth/">Read more on the RISE algorithm</a></li><li>Optimized for portrait compositions and emotional expressiveness</li></ul><h2 id="inference-configuration-miyo0a6wo">Inference Configuration</h2>Recommended Settings:<ul><li>Sampling Method: DDIM or DDIM CFG++</li><li>Sampling Steps: 30</li><li>Resolution: 512x768</li><li>CFG Scale: 12</li></ul>Prompt Prefix:<pre><code>masterpiece, high quality, 1girl, &lt;name&gt;, &lt;dress&gt;, &lt;facial expression&gt;, ...</code></pre>Negative Prompt:<pre><code>worst quality, low quality, blurry, monochrome, sketch, lineart, bad anatomy, deformed, extra limbs, missing limbs, poorly drawn face, bad proportions, ugly, disfigured, bad lighting</code></pre><h2 id="upscaling-(hires.-fix)-hc0muuhh5">Upscaling (Hires. Fix)</h2>LDSR Upscaler (Recommended):<ul><li>Denoising Strength: 0.35 - 0.75</li><li>Upscale Factor: 1.5x</li></ul>LDSR has shown the best results with this model, especially for preserving fine line art and facial features.<h2 id="results-vffchgot3">Results</h2>The model performs exceptionally well in generating charming anime girls with appealing facial expressions and balanced compositions. It demonstrates high prompt responsiveness and low failure rate even at low sampling steps.<h2 id="conclusion-2j969sv3q">Conclusion</h2>The Fallen Angel demonstrates that simplicity in generation can coexist with quality, thanks in part to the newly proposed RISE scheduler. It’s ideal for those who want efficient, high-quality anime generations without prompt fatigue.

Fallen Angel

Revision 2Codename: My Little Princess is My AngelThis is an improved model with fine-tune training dataset, as usual, I train on my puny RTX 3080 10GB, I am glad this model come-out great even training take nearly 6 hours to complete!There are additional prompt (trigger words) at my website such as <code>bekkankou</code> My next project to have model that have multiple art style, as I finished writing Anime Training Wiki (Guide)Oh, My model are not VAE bake-in, you need use AnythingV3 VAE to improve colour!

fallenAngel_rev2.ckpt

fallenAngel_rev2.yaml

Rev 2

Rev 89

Rev 35

Rev 6

Rev 5

Rev 4 (baked vae)

Rev 3

Rev 2.1-beta

Rev 1

This is a futa model built from hundreds of hours of research and training attempts. It is using fluffyrock-768-lion-low-lr-e38-offset-noise-e6 as a base. the base model this was trained on can be found in this discord server: <a target="_blank" rel="ugc" href="https://discord.gg/furrydiffusion">https://discord.gg/furrydiffusion</a>CFG goes from 4-20+, sample steps go from 20-150 on Euler A with depreciating gains after 80 steps from my tests.it's using E621 tags, not dabooru. you can get the autocomplete extension in the SD webui exstensions tab, then in the options you can switch to E621 for easy autocomplete.futanari is the keyword, but gynomorph also works (furry art generation is possible with this tag).i did my best to add all the good stuff, breast sizes go from flat chested -&gt; small breasts -&gt; medium breasts -&gt; big breasts -&gt; hyper breasts in that order. ball size follows about the same as breast size.it offers flaccid, erect, foreskin (retracted, partially retracted, unretracted). it's not great at micropenis so watch out for that.animal genitals work as well if your into that (canine/equine).you're gonna need to use an artist name as a base, i use ricegnat for a more "waifu" look. but just about any artist on E621 with more than 50 images should work fine, you can even blend them by adding multiple artists!noise offset was used, so you get great contrast!the VAE i used was vae-ft-mse-840000-ema-pruned, but you can experiment with others!in the event of too much furry bleed through make sure to add "anthro, fur, tail" in the negative prompt or human in the prompt.enjoy everyone!

futapocalypse 1.0

futapocalypse10_v10.safetensors

futapocalypse10_v10.yaml

v1.0

The model is trained based on the several fashion dress from <a target="_blank" rel="ugc" href="https://lexica.art/">https://lexica.art/</a>, you can try key word "my_dress" as one of the prompts and example "my_dress, 1girl,long_hair, makeup, solo,standing," to try. please feedback to me if any questions.

Fashion Dress

fashionDress_v10.safetensors

fashionDress_v10.yaml

A 89th training attempt, two stage training where 2nd stage using my new scheduler algorithm (RISE Inverse Stable Evolution)

fallenAngel_rev89.yaml

fallenAngel_rev89.safetensors

Trained model using PerfectDeliberate as a source. Model trained using someone very close to me.

SexyRed

sexyred_V10.yaml

sexyred_V10.safetensors

SexyRed v1.0

Revision 1 Public ReleaseCodename: The Fallen Angel Next DoorI choose "Fallen Angel" as model name because of "Angel Next Door Spoils Me Rotten" and "Kubo Won't Let Me Be Invisible". A very beautiful anime and heartwarming.It's motivate me to create beautiful angel and make two universe into one where Mahiru and Nagisa in one A.I. world. I manage to create very beautiful angel and upload to my <a rel="ugc" href="https://www.pixiv.net/en/artworks/104720959">pixiv</a>, and it has been re-upload to many web such as <a rel="ugc" href="https://www.reddit.com/r/OtonariNoTenshiSama/comments/10pi3me/mahiru_shiina_as_a_lycoris_agent_by_anime4000/">reddit</a>, <a rel="ugc" href="https://www.facebook.com/groups/237413568075701/posts/728982098918843/">facebook</a>, etc..Also I add "Saving 80,000 Gold in Another World for My Retirement" anime, Mitsuha and Sabine before finalising the model.I am glad Stable Diffusion can create beautiful inference even on sane RTX 3080 10GB.After this, adding more dataset will train longer, I wish I have better GPU to overcome this.With trick splitting the UNET and Text Encoder is very good method actually, where UNET and Text Encoder have own settings and no need to achieve 100 UNET Epoch.NOTE: Some example are using UniPC sampling method, to replicate, please update your SD WebUI. However, UniPC can be similar with DDIM at some scene.

fallenAngel_rev1.ckpt

fallenAngel_rev1.yaml

I added 3 new anime of this seasons, source is from streaming site up to 2nd Episode, screenshot video frame and properly upscale and remove H.264 artifacts. (Ripping from frames Blu-ray is recommended for up-coming SDXL)Suki na Ko ga Megane wo Wasureta<ol><li>Mie Ai</li></ol>Uchi no Kaisha no Chiisai Senpai no Hanashi<ol><li>Katase Shiori</li></ol>Yumemiru Danshi wa Genjitsushugisha<ol><li>Aizawa Rena</li><li>Ashida Kei</li><li>Inatomi Yuyu</li><li>Natsukawa Aika</li><li>Sajou Kaede</li><li>Shinomiya Rin</li></ol>School Uniform for Yumemiru Danshi wa Genjitsushugisha prompt as follow:<ol><li>kouetsu uniform (Global, random cardigan and tie/bow)</li><li>grey cardigan (aizawa rena)</li><li>black cardigan (sajou kaede)</li><li>cardigan (natsukawa aika)</li><li>red tie (natsukawa aika &amp; aizawa rena)</li><li>green tie (sajou kaede &amp; shinomiya rin)</li><li>red bow (ashida kei)</li><li>blue bow (inatomi yuyu)</li></ol>Special prompt for Mie Ai (Suki na Ko ga Megane wo Wasureta)<ol><li>glasses (Mie Ai will wearing glasses)</li><li>squinting (Mie Ai will squinting the eyes)</li><li>glasses, squinting (Mie Ai will do both, usually her in deep expression)</li></ol>Expect this checkpoint will be get better once ripping Up to 12th Episode, ripping video frame is common practice most anime checkpoint here.

fallenAngel_rev3.safetensors

fallenAngel_rev3.yaml

A more special version than the rest of the Kawai model. Hope you will like. ( •̀ ω •́ )✧

projectKKawai_preNightly130523.safetensors

projectKKawai_preNightly130523.yaml

<img src="https://imagecache.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/cb7dc224-0596-47f6-3225-493694697400/width=525/cb7dc224-0596-47f6-3225-493694697400.jpeg" /><a target="_blank" rel="ugc" href="https://ko-fi.com/N4N4KF4BO">https://ko-fi.com/N4N4KF4BO</a>The AI virtual model trained based on the prototype of Dilmurat, all the pictures appearing in the account are generated by AI software Stable Diffusion, and the characters in the pictures do not exist in reality. Any theft or reproduction of the pictures from this account for fraud or other commercial purposes is the responsibility of the infringer. 以 Dilmurat 为原型基础训练的 Ai虚拟模型，账号中所出现的所有图片均由ai软件Stable Diffusion生成，图片中人物非真实存在。任何盗用或搬运本账号图片用于诈骗或其它商业用途的，责任由盗用者本人承担。

Dilmurat_Real

***** Running training ***** Num batches each epoch = 460 Num Epochs = 150 Batch Size Per Device = 1 Gradient Accumulation steps = 1 Total train batch size (w. parallel, distributed &amp; accumulation) = 1 Text Encoder Epochs: 112 Total optimization steps = 34500 Total training steps = 69000 Resuming from checkpoint: False First resume epoch: 0 First resume step: 0 Lora: True, Optimizer: 8bit AdamW, Prec: fp16 Gradient Checkpointing: False EMA: False UNET: True Freeze CLIP Normalization Layers: False LR: 1e-05 LoRA Text Encoder LR: 0.0005

Files with SHA256: c6084a2785bbe65e2fb6e98234ee3580a00ea8d1224c6ddb1b0cae032653867f