<h1 id="heading-1863">Introduction</h1>This is a SDXL finetuned model based on <a target="_blank" rel="ugc" href="https://civarchive.com/models/136389?modelVersionId=167926">KohakuXL</a>. The original intention of creating this model is to transplant the generated effects of <a target="_blank" rel="ugc" href="https://civarchive.com/models/9139?modelVersionId=175957">YesMix</a> into the SDXL model, even though it's 'a little' difficult.<h1 id="heading-3187">Recipe</h1><ol><li>Finetune the <a target="_blank" rel="ugc" href="https://civarchive.com/models/136389?modelVersionId=167926">KohakuXL</a> using 7000 anime images randomly selected in a dataset that contains more than 10k images.</li><li>Use <a target="_blank" rel="ugc" href="https://github.com/p1atdev/LECO">LECO</a> method to weaken the negative tags and obtain a LORA, then merge this LORA in the model.</li></ol><h1 id="heading-3188">Usage</h1><ul><li>The <a target="_blank" rel="ugc" href="https://github.com/p1atdev/LECO">LECO</a> method is used in this model to weaken the negative tags, so that good results can be obtained without inputting numerous negative tags.<ul><li>The positive tags involved in training are:<code>masterpiece, best quality, extremely delicate and beautiful, highres, original</code></li><li>The negative tags involved in training are:<code>lowres, bad art, ugly, worst quality, low quality, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, jpeg artifacts, signature, watermark, username, bad feet, poorly drawn hands, poorly drawn face, mutation, mutated, deformed, extra fingers, extra limbs, missing limbs, extra legs, extra arms, malformed limbs, long neck, extra feet</code></li></ul></li></ul><h2 id="heading-3189">Generating Settings</h2><ul><li>Sampler: Euler a is recommended. Using DDIM or DPM-based samplers will generate some unexpected noise, the reason is unknown.</li><li>Resolution: It's flexible. But max width or height should be not exceed 1536.</li><li>Clip_skip: 1 or 2.</li><li>Eta noise delta: 31337 or 0</li><li>Upscale Method: HiresFix+<a target="_blank" rel="ugc" href="https://huggingface.co/konohashinobi4/4xAnimesharp/tree/main">4xAnimeSharp</a></li><li>VAE: <a target="_blank" rel="ugc" href="https://huggingface.co/stabilityai/sdxl-vae/blob/main/sdxl_vae.safetensors">SDXL VAE</a> is recommended.</li><li>Embedding: <a target="_blank" rel="ugc" href="https://civarchive.com/models/118418?modelVersionId=134583">NegativeXL</a> is recommended.</li><li>LORAs: <a target="_blank" rel="ugc" href="https://civarchive.com/models/8217/character-art-stylefashion-girl-sdxl-update">Fashion Girl</a> is supported. Other anime LORAs have not been tested so far. <a target="_blank" rel="ugc" href="https://civarchive.com/models/116480/sdxl10neg4allxlboth-positive-high-qualitydetails-and-negative-worse-qualitybad-hand-in-one-lora">Neg4All_XL</a> may be also supported.</li></ul><h1 id="heading-3190">Cautions</h1><h3 id="heading-17">The quality of NSFW images generated by this model are not as good as those generated by NAI-based model XD.</h3><ul><li>Disable 'face restoration' during generating if you do not wanna generate realistic images.</li></ul>

Files with SHA256: bee9152767182c9b6c0ef20ea26f0eee25359f49b78ad6fa0a77412a5a656cd0