<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><p><em>The Fallen Angel</em> 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—<strong>RISE (Rise Inverse Stable Evolution)</strong>—as part of ongoing research into scheduler behavior within DreamBooth. This is the <strong>89th training iteration</strong> in my series, and the resulting model performs well with simple poses, high-resolution upscaling, and minimal prompt engineering.</p><h2 id="introduction-6wnghkqdr">Introduction</h2><p>Anime generation in diffusion models often requires complex prompt crafting and intensive tuning. <em>The Fallen Angel</em> 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.</p><h2 id="methodology-kekbjy6ur">Methodology</h2><h3 id="model-focus:-8klp5ghlm">Model Focus:</h3><ul><li><p>Style: Anime (1girl, clean pose, expressive characters)</p></li><li><p>Prompt Efficiency: Reduced prompt complexity</p></li><li><p>Orientation: Primarily 512x768 (portrait)</p></li></ul><h3 id="training-details:-2o1tt5o8s">Training Details:</h3><ul><li><p><strong>Training Iteration:</strong> #89</p></li><li><p><strong>Custom Scheduler:</strong> RISE (Rise Inverse Stable Evolution)<br /><a target="_blank" rel="ugc" href="https://www.hitoha.moe/rise-algorithm-for-dreambooth/">Read more on the RISE algorithm</a></p></li><li><p>Optimized for portrait compositions and emotional expressiveness</p></li></ul><h2 id="inference-configuration-miyo0a6wo">Inference Configuration</h2><p><strong>Recommended Settings:</strong></p><ul><li><p><strong>Sampling Method:</strong> DDIM or DDIM CFG++</p></li><li><p><strong>Sampling Steps:</strong> 30</p></li><li><p><strong>Resolution:</strong> 512x768</p></li><li><p><strong>CFG Scale:</strong> 12</p></li></ul><p><strong>Prompt Prefix:</strong></p><pre><code>masterpiece, high quality, 1girl, &lt;name&gt;, &lt;dress&gt;, &lt;facial expression&gt;, ...</code></pre><p><strong>Negative Prompt:</strong></p><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><p><strong>LDSR Upscaler (Recommended):</strong></p><ul><li><p><strong>Denoising Strength:</strong> 0.35 - 0.75</p></li><li><p><strong>Upscale Factor:</strong> 1.5x</p></li></ul><p>LDSR has shown the best results with this model, especially for preserving fine line art and facial features.</p><h2 id="results-vffchgot3">Results</h2><p>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.</p><h2 id="conclusion-2j969sv3q">Conclusion</h2><p><em>The Fallen Angel</em> 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.</p>

Fallen Angel

<p>A huge leap from Revision 6 to 35, multiple training attempt in past 3 month to get model right and produce better results with simple prompt. My dataset with manual tag clean-up and tag consolidation to ensure better inference.</p><p>As always, with latest my favorite waifu :D</p>

Files with SHA256: bc959311230915d7c529fd734536f62e1749f3a2c09ce4583641c9eaee0de4b8