💬 Discord: <a target="_blank" rel="ugc" href="https://discord.gg/reMJwfNp3j">join here</a> · ☕ Ko-fi: <a target="_blank" rel="ugc" href="https://ko-fi.com/techannel">support the factory</a><h1 id="halation-an-anima-finetune">Halation, an Anima finetune</h1>The bloom around the light. v0.1 was the gentle pass. v0.2 is the one that got pushed.<hr />Full finetune of <a target="_blank" rel="ugc" href="https://civarchive.com/models/2458426">Anima</a>. Not a merge, not a LoRA extract. Every tensor in the file moved.What I want from Halation has not changed: light with some dimension to it, and materials that read as materials. Anima's vocabulary is untouched, same tags and same characters. Only the default look changes.v0.1 promised the next one would get pushed harder. This is that one. The quality ladder went in properly this time, so the score and rating tokens below are trained vocabulary, not decoration.<h2 id="versions">🧩 Versions</h2><ul><li>🔵 v0.2, a full checkpoint for Anima. The deep pass. No trigger word and no weight slider, because it is the model. You still need the Anima text encoder and VAE, like any Anima checkpoint.</li><li>⚪ v0.1, the first, gentle pass. Kept for comparison; its own settings below still apply to it.</li></ul><h2 id="prompt-v0.2">📋 Prompt — v0.2</h2>Positive prefix. The full ladder, and keep a rating token in it:<pre><code>masterpiece, best quality, high quality, normal quality, score_7, score_9, sensitive, ...</code></pre>Negative:<pre><code>low quality, worst quality, score_1, score_2, score_3, bad hands, missing fingers, bad anatomy, blurry, fused limbs, artist name</code></pre>Settings behind every v0.2 image: 50 steps, CFG 4.0, <code>euler_ancestral</code> with <code>normal</code>, 1024x1536, CLIP skip 1.<blockquote>v0.2 showcase images ran the house face-detail second pass (a light 0.26-denoise face refine). The base render is one pass at 50 steps.</blockquote><h2 id="prompt-v0.1">📋 Prompt — v0.1</h2>Positive prefix. The rating token is the part not to skip:<pre><code>masterpiece, best quality, score_7, score_9, safe, ...</code></pre>Negative. This does more for the image than the score tokens:<pre><code>worst quality, low quality, score_1, score_2, score_3, artist name, blurry, shiny skin</code></pre>Settings behind every v0.1 image: 30 to 50 steps, CFG 4.0, <code>euler_ancestral</code> with <code>simple</code>, 1024x1536, CLIP skip 1.<blockquote>Every v0.1 image is raw output. No face detailer, no upscale, no inpainting. One pass at 50 steps.</blockquote><h2 id="tips">💡 Tips</h2><ul><li>Always put a rating token in. <code>safe</code>, <code>sensitive</code>, <code>nsfw</code> or <code>explicit</code>. Every training caption carried one, so a prompt without it is off-distribution and the model fills the slot from its prior — and that prior leans <code>sensitive</code>, so innocuous prompts drift. Say what you want and it listens.</li><li>The negative is worth more than the score tokens. The bare prompt gives you a finished image, but hazy. The negative on its own jumps richness and contrast hard. Score tokens polish what is already there.</li><li>Those negative words are real levers, not decoration. They were tagged explicitly in training instead of being absorbed as the house look, so negating them takes them away.</li><li><code>score_9</code> is the grounding dial. It pulls toward realism and weight. <code>score_7, score_9</code> together land somewhere softer. Pick by taste.</li><li>It still knows characters. Plain-prompt franchise characters come out right, no LoRA — Anima's knowledge survived both passes intact.</li></ul><hr />More of my models: <a target="_blank" rel="ugc" href="https://civarchive.com/user/Techannel">Techannel on Civitai</a>. Likes and Buzz keep the factory running ❤️

Files with SHA256: 4da838401fa98c9ecbaf79ee260bd7dc5f0469ec69cd8d2ddfa1c95ceab05df3