<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 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.

Files with SHA256: 9eaaf24f98e82f298aae113424d83c0c30776c4df323107fcbfefc69758a1701