<h1 id="anidream-trueanime">AniDream TrueAnime</h1><blockquote>Drawn, not rendered. No plastic. No oil. Just anime.Latex, metal, water and glass still shine — only skin stays matte.</blockquote><hr /><h2 id="quick-start">⚡ Quick Start</h2>For regular Anima (aka 2B)<ul><li>Sampler: er_sde · Scheduler: beta</li><li>Steps: 30–50 · CFG: 5–6 · Sigma Shift: 3–7</li><li>Resolution: up to 1536² natively, higher with Highres Boost-2K</li></ul>For Anima 2.9B<ul><li>Sampler: res_multistep, er_sde, euler, euler a · Scheduler: sgm_uniform, beta, beta57, linear_quadratic</li><li>Steps: 28–50 · CFG: 3.5–5 · Sigma Shift: 5–7, 6 as default. Go higher (8–9) for big canvases and busy compositions, lower for more fine detail.</li><li>Resolution: up to 1536² natively, higher with Highres Boost-2K (2B LoRAs work through ComfyUI-Anima-2.9B-loraPatch)</li><li>Text encoder: AnimaTextEncoderQwen30_fp32CreatorIwakuReal (recommended, attached in additional files), base <code>qwen_3_06b</code> works almost the same.</li><li>VAE: Qwen-Image VAE (standard, also attached in additional files) or any Qwen-Image-compatible VAE. </li></ul>Sigma Shift: lower (3–4) = more fine detail, but weaker composition; higher (6–7) = stronger composition. On big canvases (~1536×1920) stay at 5–7 to avoid duplicates.Multi-character scenes: start with Shift 7–9 — best hands and composition. If a character loses a small feature (tail, accessory), try Shift 3 — or just try another seed: the seed often matters more than the Shift. Shift 5 is not always the safe middle. To fix a small detail without changing the composition, use a variation seed with low strength (≈0.1–0.3) instead of a new seed.Extra duplicates (two phones, extra props) can be removed with <code>multiple X</code> in the negative, e.g. <code>multiple phones, multiple cellphones</code>.That's it. Generate.<hr /><h2 id="several-characters-in-one-image">👥 Several Characters in One Image</h2><ul><li>Give each character its own block with a weight: <code>&lt;weight[1.3]:...&gt;</code> (ComfyUI / Forge: <code>(...:1.3)</code>). The right weight depends on the sampler: er_sde ≈ 1.3–1.4, euler ≈ 1.4. Above 1.5 things start to break.</li><li>Repeat the character's name in every sentence: <code>Miku has...</code>, <code>Miku is saluting.</code> A sentence without a name can jump to a neighbor.</li><li>Write the blocks in reading order (left → right, top → bottom), the same as on the canvas.</li><li>Different outfits = fewer leaks. Same clothes on everyone make colors jump between characters.</li><li>Keep backgrounds plain. Patterns (flowers, charts) can merge panels into one.</li><li>Small details (nail polish) leak the most. Skip them if you don't need them.</li><li>Check your brackets! One missing <code>]</code> and the character's features scatter to everyone else.</li></ul><hr /><h2 id="prompt">✍️ Prompt</h2>Order matters: <code>quality tags</code> → <code>character, outfit, pose, scene</code> → <code>style / artists</code> → <code>detail tags</code>Style tags go at the end (right before the postfix), not at the start, or they leak into the scene.Tags for WHAT, short sentences for HOW: <code>1girl, frieren, white bodysuit</code> + <code>She leans on the motorcycle, looking over her shoulder.</code>Punctuation matters: separate tags with commas, but end sentences with a period. The period tells the model where a sentence ends, so the following tags don't "stick" to it.<a target="_blank" rel="ugc" href="https://danbooru.donmai.us/wiki_pages/tag_group:subjective">Subjective tags</a> work: <code>beautiful, cute, handsome, gorgeous, sexy</code> and the like actually change the result, so use them to set the mood of a character.Prefix:<pre><code>masterpiece, best quality, score_9, score_8, score_7, newest, year 2025, highres, absurdres, very aesthetic, ultra detailed,
</code></pre>Postfix:<pre><code>anatomically perfect, correct proportions, natural posture, ultra detailed face, ultra detailed eyes, ultra detailed hair, ultra detailed hair strands, ultra detailed clothes, ultra-detailed background, detailed scenery,
</code></pre>Optional: for more dynamic, less "statue-like" poses<pre><code>asymmetrical pose, asymmetrical hand pose,
</code></pre>Negative:<pre><code>worst quality, low quality, early, old, score_1, score_2, score_3, cartoon, graphic, painting, crayon, graphite, abstract, glitch, deformed, mutated, ugly, disfigured, long body, bad anatomy, bad hands, missing fingers, extra fingers, extra digits, fewer digits, cropped, very displeasing, artist name, blurry, jpeg artifacts, lowres, censor,</code></pre>Optional (extra negative for cleaner results):<pre><code>Poorly drawn anatomy, especially in the hands, feet, face, and joints.
monochrome, greyscale, spot color, muted color, pale color, desaturated,
realistic teeth,
border, white border, black border, picture frame, framed, letterboxed, pillarboxed, windowboxed, polaroid border, film frame, photo border, edge markings, vignette,
bad text, gibberish text,
surreal, photo background, 3d, realistic, photorealistic, skin texture, volumetric lighting, blender,
Bad perspective, aspect ratio, reflections, and shading.
Clipping, misalignment, and warped or disjointed geometry.
minimalism, low detail background, absurdly detailed composition,
engineering nonsense, cartoon logic, what, bad gun anatomy, bad vehicle anatomy, bad food,
bad clothes, impossible clothes,
symmetrical pose, symmetrical hand pose, symmetry,</code></pre><pre><code>Use (bad anatomy, bad hands, ...:1.2) if anatomy looks weird.</code></pre><hr /><h2 id="text-on-images">🔤 Text on Images</h2>The model can write text, but it's not a typewriter. What we found in testing:<ul><li>Put the text in quotes, say where it goes, end with a period: <code>One line of text in the top left corner: "Welcome to the future!".</code></li><li>Give it a small weight (~1.2). Without it the model often skips the text; more than 1.2 doesn't make it more accurate. SwarmUI: <code>&lt;weight[1.2]:...&gt;</code> · ComfyUI / Forge: <code>(...:1.2)</code></li><li>Short, common words work best. Long or rare words (like "distant") may come out misspelled no matter what.</li><li>One line per corner works better than two lines in one place. Print fonts are easier for the model than cursive.</li><li>Lower Sigma Shift (3–4) seems to help small details like letters.</li><li>Don't double the period: <code>"way..."</code> + <code>.</code> outside the quotes = extra dots in the image.</li></ul>If one word is still wrong: inpaint only the text with a mask (Creativity 0.6–0.8), or just fix a letter or two in an image editor. It's often faster.<hr /><h2 id="samplers">🎛️ Samplers</h2><ul><li>er_sde — clean &amp; safe (default)</li><li>euler — softer, slightly more creative</li><li>euler_a — thin soft lines; good for complex scenes with several characters</li><li>dpmpp_2m_sde_gpu — more dynamic, but has a mind of its own</li><li>dpmpp_2m — razor sharp</li><li>dpmpp_3m_sde — not for generating from scratch; good for img2img</li></ul><hr /><h2 id="upscaling">🔍 Upscaling</h2>Skip Hires Fix — it tends to distort the image.Use Img2Img / Init Image instead:<ol><li>Generate the image as usual.</li><li>Send it to Img2Img / Init Image with low denoise (Init Image Creativity 0.25–0.35, or higher if you're experimenting / inpainting).</li></ol>Sigma Shift for Img2Img:<ul><li>Low denoise (≈0.25–0.35): lower the Shift (≈3–4) — the composition is already there, let the model spend its steps on detail.</li><li>High denoise (0.5+): keep the same Shift as the first pass, or a middle value (5).</li></ul>The composition stays yours — the model only adds detail.<hr /><h2 id="known-limits">⚠️ Known Limits</h2>Anatomy is stable, but it's still a diffusion model. Occasional "mutations" (an extra finger, a doubled braid, a third leg) can appear within normal statistical error. They depend on settings, seed, resolution and LoRAs.If it happens:<ul><li>try another seed;</li><li>try a different sampler and/or scheduler (separately or together);</li><li>lower the resolution or raise Sigma Shift;</li><li>reduce the number or strength of LoRAs;</li><li>add <a target="_blank" rel="ugc" href="https://civarchive.com/models/2832931/negative-anatomy-fixer-detailer">Negative Anatomy Fixer</a> at a negative weight (around −0.6…−1). It helps, but it's not a 100% cure.</li></ul><hr /><h2 id="recommended-loras">✅ Recommended LoRAs</h2><ul><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/2673546?modelVersionId=3001940">Highres Boost-2K</a> (1.0) — for big canvases</li><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/929497/aesthetic-quality-modifiers-masterpiece?modelVersionId=2961717">anima-base-1-masterpiece</a> (1.0) — overall quality</li><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/2738621/2d-3d-slider?modelVersionId=3079658">2D-3D slider</a> (−1…−2) — highly recommended for even more 2D</li><li><a target="_blank" rel="ugc" href="https://civarchive.com/models/2684020/better-text-rendering?modelVersionId=3013765">Better Text Rendering</a> (0.6–0.8) — if you need text (helps, but isn't magic)</li></ul><hr /><h2 id="don't">❌ Don't</h2><ul><li>Don't add Reflection or anti-oil sliders. They're already baked in — a double dose burns the image.</li><li>Don't push CFG above 6. Burned colors, harsh lines.</li><li>Don't stack LoRAs mindlessly. Sometimes less is better.</li><li>Anima is not SDXL. No tag walls, no BREAK, and no SDXL LoRAs, embeddings, etc. It's like putting diesel in a petrol car: both are fuel, both fit the same tank — and it still wrecks the engine.</li></ul>or...Prepare for unforeseen consequences.<hr /><h2 id="thanks">🙏 Thanks</h2><ul><li><a target="_blank" rel="ugc" href="https://civarchive.com/user/pesy">pesy</a> — for <a target="_blank" rel="ugc" href="https://civarchive.com/models/2675133/anima-reflection?modelVersionId=3003728">Anima-Reflection⬇</a>, the LoRA that made the matte skin possible.</li></ul><hr />Tests and ideas by the author. English isn't my native language, so the text was written and edited with the help of Claude.

AniDream TrueAnime

TrueAnime29_v1.safetensors

Files with SHA256: db285a9f8cbb7ee8ef8b0ef4d7f33a94bd6b845c938d3884aace54bfc8fb0ee3