<h3 id="realcosplay-is-a-multi-style-model-not-limited-to-realism-alone.">RealCosplay is a multi-style model, not limited to realism alone.</h3>⛔ DON'T USE DPM++ 2M or High Cfg in 1536px verison!! They will burn your image and add strange noise<hr /><h2 id="recommended-settings">Recommended settings：</h2><ul><li>CFG: 3~3.5</li><li>Steps: 28 ~ 40</li><li>Sampling Method: [Euler a✅] and [DPM++ 2S a✅] have been verified (⚠️Others are likely to generate color noise or strange patterns. DON'T USE <s>DPM++ 2M</s><s>⚠️</s>, <s>Euler</s><s>⚠️</s>, etc.They may not support illust2.0 1536px model)</li><li>Resolution: Train on 1536px. supports 512~1536 resolutions . In my test, 1024px has better logic and adherence. 1536px has finer texture details.You can generate at 1024px first, then Hires.fix t o 1536px. Total pixel count within 2.35 million is acceptable, for example, the resolution of 1824x1248 also very stable.</li><li>postive prompts:photograph,cosplay,illustration,</li><li>negative prompts for photographillustration,anime,sketch,comic,cartoon,toon,oil painting \(medium\),flat color,outline,3D,2.5D,2D,unreal,unity,blender,anime,smooth skin,semi-anthro,impasto,</li></ul><h3 id="prompt-guidelinestrained-trigger">Prompt Guidelines（trained trigger）</h3><ul><li>Real style：photograph,cosplay,</li><li>Anime style：illustration,anime,</li><li>Ethnicity: european, asian,</li><li>tip：The prompt "realistic" will generate semi-realistic images, not photo-realistic images.</li></ul><h3 id="safety-guidelinessame-as-illustxl-2.0">Safety Guidelines（same as IllustXL 2.0）：</h3><ul><li>Prompt：adult content, r-18, nsfw, explicit, sfw, safe for work</li></ul>

RealCosplay 1536px

<ol><li>Full finetuning from Illustrious XL 2.0</li><li>trained on 7000+ images and Tag+NLP mixed caption</li></ol>

realcosplay1536px_v21.safetensors

v2.1

RealCosplay is a multi-style model, not limited to realism alone.If you like this model, you can support my work on <a href="https://app.unifans.io/c/sennke" rel="ugc noopener noreferrer nofollow" target="_blank">Unifans</a>.Or give my model a ❤️Unifans: <a href="https://app.unifans.io/c/sennke" rel="ugc noopener noreferrer nofollow" target="_blank">https://app.unifans.io/c/sennke</a> (sample workflow, more sample image)X: <a href="https://x.com/SennkeEknn" rel="ugc noopener noreferrer nofollow" target="_blank">https://x.com/SennkeEknn</a>------------------------------------------------------------------------------------------------1536 vs 1024 【The default style is still anime style】 ✅Better small face quality 更好的小脸质量✅Better skin detail 更好的皮肤细节✅Better understanding of hand movements 更好的手部理解⚠️Style diversity is slightly weaker at high resolution 高分辨率时风格化较弱⚠️30+ steps is better, but the speed will become slower 需要更高步数Overall, 1536px version is my favorite. I may take a break from it for a while.❤️------------------------------------------------------------------------------------------------Recommended settings：<ul><li>CFG: 3~3.5</li><li>Steps: 28 ~ 40</li><li>Sampling Method: [Euler a✅] and [DPM++ 2S a✅] have been verified (⚠️Others are likely to generate color noise or strange patterns. They may not support illust2.0 1536px model, such as <s>DPM++ 2M⚠️</s>, <s>Euler⚠️</s>, etc.)</li><li>Resolution: Train on 1536px. supports 512~1536 resolutions . In my test, 1024px has better logic and adherence. 1536px has finer texture <a href="http://details.You" rel="ugc noopener noreferrer nofollow" target="_blank">details.You</a> can generate at 1024px first, then Hires.fix to 1536px. Total pixel count within 2.35 million is acceptable, for example, the resolution of 1824x1248 also very stable.</li><li>postive prompts:photograph,cosplay,illustration,</li><li>negative prompts for photographillustration,anime,sketch,comic,cartoon,toon,oil painting \(medium\),flat color,outline,3D,2.5D,2D,unreal,unity,blender,anime,smooth skin,semi-anthro,impasto,</li></ul>------------------------------------------------------------------------------------------------Prompt Guidelines（trained trigger）<ul><li>Real style：photograph,cosplay,</li><li>Anime style：illustration,anime,</li><li>Ethnicity: european, asian,</li><li>tip：The prompt "realistic" will generate semi-realistic images, not photo-realistic images.-----------------------------------------------------------------------------------------------</li></ul>Safety Guidelines（same as IllustXL 2.0）：<ul><li>Prompt：***** content, r-18, nsfw, explicit, sfw, safe for work</li></ul>

Full finetuning from Illustrious XL 2.0

trained on 7000+ images and Tag+NLP mixed caption

Files with SHA256: b2e1132553069c7b1a5a48bd0cec87dbc73192e2837be0432261e4bdccd9ead2