<h1>⭐ StarIllustriousTurbo</h1>
<p><em>Starlight, soft glow, and a sky full of wishes, in just 12 steps.</em></p>
<p>StarIllustriousTurbo is <strong>StarIllustrious</strong> with Herrscher_AGGA's <strong>DMD2 V7.5 (BTN)</strong> distillation LoRA baked in. It keeps StarIllustrious's deep night-blue palettes, glowing highlights, clean faces and expressive, sparkling eyes, but finishes an image in <strong>12 steps at CFG 1</strong> instead of 30 steps at CFG 5. In our tests that makes each image about <strong>four times faster</strong>. DMD2 also adds a little extra contrast and crisp detail.</p>
<p>It shines on portraits and upper-body shots, starry skies, glowing magic and anything that catches the light.</p>
<p><strong>No trigger words.</strong> Prompt with Danbooru tags.</p>
<hr />
<h2>🌠 The recipe</h2>
<ol>
<li><strong>Base:</strong> <a rel="ugc" href="https://civarchive.com/models/3017804">StarIllustrious</a>, a SLERP merge (t = 0.45) of <a rel="ugc" href="https://civarchive.com/models/2430439">WaiSnackMix Collection v1.1</a> by MGHerder and <a rel="ugc" href="https://civarchive.com/models/2961356">Luminelle v1.0</a> by VIA_AI</li>
<li><strong>Baked in:</strong> <a rel="ugc" href="https://civarchive.com/models/1850983?modelVersionId=2935186">DMD2 | Fewer Steps V7.5 | BTN</a> by Herrscher_AGGA, at strength 1.0</li>
</ol>
<p>DMD2 is a distillation LoRA: it teaches the model to finish an image in a few steps without classifier-free guidance. It was folded into the UNet and both text encoders, so there is nothing extra to load. The VAE is StarIllustrious's, unchanged. Merged and converted locally with stable-diffusion.cpp.</p>
<h2>✨ Settings</h2>
<ul>
<li><strong>Steps:</strong> 12. Anything from 8 to 16 works.</li>
<li><strong>CFG:</strong> 1, up to about 1.5.</li>
<li><strong>Sampler / scheduler:</strong> Euler a + Karras</li>
<li><strong>Resolution:</strong> 832×1216 portrait, 1216×832 landscape, 1024×1024 square</li>
<li><strong>Negative prompt:</strong> has no effect at CFG 1. Above CFG 1, use <code>worst quality, low quality, lowres, bad anatomy, bad hands, extra fingers, jpeg artifacts, blurry, artist name, signature, watermark</code></li>
<li><strong>VAE:</strong> baked in</li>
<li><strong>No extra speed LoRA:</strong> DMD2 is already built in, so don't stack another DMD2, LCM, Lightning or Hyper LoRA on top.</li>
</ul>
<h2>🔭 Prompting</h2>
<p>StarIllustriousTurbo reads Danbooru tags, not sentences:</p>
<ul>
<li>Comma-separated tags, most important first: <code>1girl</code> / <code>1boy</code> / <code>no humans</code>, then hair, eyes, outfit, expression and pose, then the setting, lighting and framing.</li>
<li>Put the quality tags at the end: <code>masterpiece, best quality, absurdres</code>. Pony-style score tags do nothing here.</li>
<li>Framing tags such as <code>portrait</code>, <code>close-up</code>, <code>upper body</code> and <code>cowboy shot</code> work well; it is at its best close in.</li>
<li>Tags whose names contain brackets need them escaped, like <code>star \(symbol\)</code>, or they are read as weights.</li>
<li>Negative prompts do nothing at CFG 1, so steer with the positive prompt. For mature content, put <code>mature female</code> in the positive; if you also want a negative such as <code>child, loli, aged down</code>, raise the CFG to 1.5.</li>
</ul>
<p>More tips in <a rel="ugc" href="https://civarchive.com/articles/23210">Arctenox's simple prompt guide for Illustrious</a>.</p>
<h2>🪐 Which file?</h2>
<ul>
<li><strong>FP16</strong> · <code>starillustriousturbo-v1.0-fp16.safetensors</code>: the full checkpoint (UNet, text encoders and VAE). Every showcase image was rendered with this file.</li>
<li><strong>FP8 (scaled)</strong> · <code>starillustriousturbo-v1.0-fp8_scaled.safetensors</code>: a lighter checkpoint for ComfyUI.</li>
<li><strong>Q8_0 / Q5_K / Q4_K</strong> · <code>starillustriousturbo-v1.0-q8_0.gguf</code> and friends: all-in-one GGUFs for stable-diffusion.cpp. ComfyUI-GGUF's UNet loader reads only the UNet from them, so take the text encoders and VAE from the FP16 file (the DMD2 LoRA also touched the text encoders).</li>
<li><strong>INT8</strong> · <code>starillustriousturbo-v1.0-int8.safetensors</code>: for ComfyUI INT8 (convrot) workflows. It also runs natively in recent stable-diffusion.cpp builds.</li>
<li><strong>W4A4</strong> · <code>starillustriousturbo-v1.0-int4_convrot.safetensors</code>: ComfyUI INT4 W4A4. Experimental, and it does not run in stable-diffusion.cpp yet.</li>
</ul>
<p>Every file is converted directly from the FP16 master, never from another quant.</p>
<h2>🌌 Content note</h2>
<p>Like StarIllustrious, StarIllustriousTurbo handles mature content when you prompt for it with <code>nsfw</code>. The last two showcase images contain nudity; everything else is SFW. All characters are adults. Please follow CivitAI's rules and the laws where you live.</p>
<h2>☄️ Cover prompt</h2>
<blockquote>
<p>1girl, solo, long hair, silver hair, floating hair, blue eyes, star-shaped pupils, star hair ornament, white dress, looking at viewer, close up, parted lips, night sky, starry sky, milky way, shooting star, light particles, wind, upper body, from below, masterpiece, best quality, absurdres</p>
</blockquote>
<p>12 steps · CFG 1 · Euler a / Karras · 832×1216 · seed 695841494</p>
<p>Every showcase image uses the same prompt as the matching image on StarIllustrious's page, and all but one (the stargazing couple) use the same seed, so you can compare the two models side by side. Each image keeps its full prompt and settings, so you can remix any of them.</p>
<h2>💫 Credits</h2>
<p>Thank you to <strong>Herrscher_AGGA</strong> (DMD2), <strong>MGHerder</strong> (WaiSnackMix Collection), <strong>VIA_AI</strong> (Luminelle), and <strong>OnomaAI</strong> for Illustrious XL. StarIllustriousTurbo would not exist without their work. 💙</p>

StarIllustriousTurbo

<p><strong>v1.0: shooting star</strong> ☄️</p>
<ul>
<li>StarIllustrious v1.0 with DMD2 V7.5 | BTN (Herrscher_AGGA) baked in at strength 1.0, in the UNet and both text encoders</li>
<li>Recommended: 12 steps · CFG 1 · Euler a / Karras · 832×1216</li>
<li>Seven formats: FP16, FP8 (scaled), Q8_0, Q5_K, Q4_K, INT8 and W4A4 (INT4 convrot), all converted from the FP16 master</li>
<li>VAE baked in</li>
</ul>

Files with SHA256: 25ca925deb13517fb9b40a89fefb04eb6f8cd9c9e39b9e07235ff8f07fb6412d