<h1 id="anima-+-controlnet-lllite-+-4k">Anima 三次采样 + ControlNet-LLLite + 4K 高清放大重绘工作流</h1><hr /><h2 id="overview">Overview 概述</h2>English: A complete production pipeline for the Anima base model, built to solve Anima's two core weaknesses — soft detail and structural drift at high resolution. The workflow fuses three Anima checkpoints, applies two LoRA stacks, runs three sequential sampling passes with ControlNet-LLLite structural control, then performs a 4K Ultimate SD Upscale with a dedicated upscale model, finishing with RTX denoise/deblur and anime color grading. Designed for clean, high-detail, structurally-stable anime output.中文： 一套针对 Anima 底模的完整生产管线，专为解决 Anima 两大核心短板——细节偏软、高分辨率下结构漂移。工作流融合三个 Anima 底模、叠加两组 LoRA 栈、串联三次采样并配合 ControlNet-LLLite 结构控制，最后进行4K Ultimate SD Upscale 超分（专用放大模型），以 RTX 降噪/去模糊与动漫调色收尾。产出干净、高细节、结构稳定的动漫成品。<hr /><h2 id="architecture-+-4k">Architecture 架构（三段式采样 + 4K 超分）</h2>English:<ul><li>Stage 1 — Base Generation（一采）: 40 steps, CFG 5, denoise 1.0, at 864×1536 (portrait) or 1536×864 (landscape).</li><li>Stage 2 — Latent Upscale &amp; Redraw（二采）: bicubic latent upscale ×1.5, then 40 steps at denoise 0.8.</li><li>Stage 3 — ControlNet-LLLite Guided Redraw（三采）: 40 steps at denoise 0.8, guided by depth/lineart LLLite to lock structure and sharpen detail.</li><li>Stage 4 — 4K Ultimate SD Upscale（4K超分）: 2× upscale with <code>4x-AnimeSharp</code> model, 30 steps at denoise 0.05, tile 2304×1296 (full-image mode) for a clean high-res finish.</li><li>Post-processing: RTX Denoise → RTX Deblur → EasyColorCorrection (Anime Moody).</li></ul>中文：<ul><li>第一采——基础生成： 40 步、CFG 5、denoise 1.0，在 864×1536（竖版）或 1536×864（横版）生成初始图。</li><li>第二采——潜空间放大重绘： bicubic 潜空间放大 1.5 倍，再以 denoise 0.8 采样 40 步。</li><li>第三采——ControlNet-LLLite 引导重绘： denoise 0.8 采样 40 步，深度/线稿 LLLite 控制锁定结构、锐化细节。</li><li>第四段——4K Ultimate SD Upscale： <code>4x-AnimeSharp</code> 放大模型 2 倍超分，30 步、denoise 0.05、tile 2304×1296（整图模式），产出干净高分辨率成片。</li><li>后处理： RTX 降噪 → RTX 去模糊 → EasyColorCorrection（Anime Moody 预设）。</li></ul><hr /><h2 id="key-features">Key Features 核心特性</h2>English:<ol><li>Triple-checkpoint fusion — chosenMixAnima_v10 + anima_aestheticV11 + miaomiaoHarem_anima16, merged via ModelMergeSimple (0.6) + ModelMergeBlocks (input 0.9 / middle 0.2 / out 0.7) for layered control of structure, style and detail.</li><li>Dual LoRA stacks — quality stack (masterpiece 0.9 + highres-aesthetic-boost 0.9 + rdbt 0.9 + light-concepts 0.6) + detail stack (rendering_detailer −0.3 + context_detailer 0.3).</li><li>ControlNet-LLLite structural control — depth + lineart LLLite keep anatomy stable across all sampling passes.</li><li>ReferenceLatent anchoring — preserves the generated image as a structural anchor during redraw.</li><li>AuraFlow shift=3 — compensates Anima's native softness for sharper lines and richer texture.</li><li>4K Ultimate SD Upscale — dedicated <code>4x-AnimeSharp</code> model + full-image tile mode for clean 2× upscale.</li><li>Camera control node — azimuth/elevation/distance-driven camera prompt injection.</li><li>RTX + color grading post-chain — cleans artifacts and applies an Anime Moody grade.</li></ol>中文：<ol><li>三底模融合——chosenMixAnima_v10 + anima_aestheticV11 + miaomiaoHarem_anima16，经 ModelMergeSimple(0.6) + ModelMergeBlocks(input 0.9/middle 0.2/out 0.7) 分层控制结构、风格与细节。</li><li>双 LoRA 栈——质量栈（masterpiece 0.9 + highres-aesthetic-boost 0.9 + rdbt 0.9 + light-concepts 0.6）+ 细节栈（rendering_detailer −0.3 + context_detailer 0.3）。</li><li>ControlNet-LLLite 结构控制——深度 + 线稿 LLLite 全程锁定解剖结构。</li><li>ReferenceLatent 锚定——重绘时以已生成图为结构锚，防止漂移。</li><li>AuraFlow shift=3——补偿 Anima 原生偏软，线条更锐、纹理更实。</li><li>4K Ultimate SD Upscale——专用 <code>4x-AnimeSharp</code> 模型 + 整图 tile 模式，干净 2 倍超分。</li><li>相机控制节点——方位角/俯仰角/距离驱动的相机提示词注入。</li><li>RTX + 调色后处理链——清理伪影并套用 Anime Moody 动漫调色。</li></ol><hr /><h2 id="model-and-lora-dependencies">Model &amp; LoRA Dependencies 依赖清单</h2>Type 类型File 文件Weight 权重CheckpointchosenMixAnima_v10.safetensors(merged)Checkpointanima_aestheticV11.safetensors(merged)CheckpointmiaomiaoHarem_anima16.safetensors(merged)LoRAanima-base-1-masterpiece-v510.9LoRAanima-highres-aesthetic-boost0.9LoRArdbt_v2.1_base_anima_b10.9LoRAanima-base-1-light-concepts-v310.6LoRArendering_detailer_base10−0.3LoRAanima_context_detailer_base100.3CLIPmiaomiaoHarem_anima16_txt.safetensors—VAEqwen_image_vae.safetensors—Upscale Model4x-AnimeSharp.safetensors—LLLiteanima-lllite-depth-10.8LLLitecontrolnetLlliteAnima_lineart0.8<hr /><h2 id="settings">Settings 参数配置</h2>English:<ul><li>Sampler: <code>euler_ancestral</code> / Scheduler: <code>normal</code></li><li>Steps: <code>40</code> (UltimateSDUpscale: <code>30</code>) / CFG: <code>5</code></li><li>Denoise: Stage1 <code>1.0</code> → Stage2 <code>0.8</code> → Stage3 <code>0.8</code> → 4K pass <code>0.05</code></li><li>Resolution: <code>864×1536</code> portrait / <code>1536×864</code> landscape</li><li>Latent upscale: <code>bicubic ×1.5</code></li><li>UltimateSDUpscale: <code>2×</code> / tile <code>2304×1296</code> / padding <code>128</code></li><li>ModelMergeSimple: <code>0.6</code> / ModelMergeBlocks: <code>input 0.9, middle 0.2, out 0.7</code></li><li>ModelSamplingAuraFlow: <code>shift=3</code></li><li>Seed: fixed, with increment mode available</li></ul>中文：<ul><li>采样器：<code>euler_ancestral</code> / 调度器：<code>normal</code></li><li>步数：<code>40</code>（UltimateSDUpscale 用 <code>30</code>）/ CFG：<code>5</code></li><li>Denoise：一采 <code>1.0</code> → 二采 <code>0.8</code> → 三采 <code>0.8</code> → 4K 超分 <code>0.05</code></li><li>分辨率：<code>864×1536</code> 竖版 / <code>1536×864</code> 横版</li><li>潜空间放大：<code>bicubic ×1.5</code></li><li>UltimateSDUpscale：<code>2×</code> / tile <code>2304×1296</code> / padding <code>128</code></li><li>ModelMergeSimple：<code>0.6</code> / ModelMergeBlocks：<code>input 0.9, middle 0.2, out 0.7</code></li><li>ModelSamplingAuraFlow：<code>shift=3</code></li><li>种子：固定，支持递增模式</li></ul><hr /><h2 id="how-to-use">How to Use 使用步骤</h2>English:<ol><li>Install required custom node packs (rgthree, KJ nodes, LLLite sd-scripts, RTX nodes, EasyColorCorrection, Layer Utility, Ultimate SD Upscale).</li><li>Download all checkpoints/LoRA/LLLite/upscale model files to correct folders.</li><li>Load the workflow and fix missing model paths.</li><li>Write your prompt (English tags + natural language works best with Anima's Qwen encoder).</li><li>Run. Adjust LoRA weights, merge ratios and denoise as needed.</li></ol>中文：<ol><li>安装所需自定义节点包（rgthree、KJ nodes、LLLite sd-scripts、RTX 节点、EasyColorCorrection、Layer Utility、Ultimate SD Upscale）。</li><li>下载全部底模/LoRA/LLLite/放大模型文件到对应文件夹。</li><li>加载工作流并修正缺失路径。</li><li>编写提示词（Anima 的 Qwen 编码器最适合英文标签 + 自然语言）。</li><li>运行，按需调整 LoRA 权重、融合比例和 denoise。</li></ol>

Files with SHA256: 882ff5b2c4ea2106966cd1204e5c02ed4963a852c39d3d224275741cbcf47332