[{"data":1,"prerenderedAt":470},["ShallowReactive",2],{"content-/zh/advanced-tutorial/ltx-workflow":3},{"id":4,"title":5,"body":6,"description":16,"extension":463,"meta":464,"navigation":465,"path":466,"seo":467,"stem":468,"__hash__":469},"content/zh/advanced-tutorial/ltx-workflow.md","精通 ComfyUI LTX 2.3：高保真视频生成实用指南",{"type":7,"value":8,"toc":444},"minimark",[9,13,17,29,35,38,43,46,50,83,90,92,96,99,102,105,130,132,136,139,143,161,171,175,184,193,202,206,213,216,236,240,247,301,305,318,329,341,348,350,354,357,360,391,393,397,400,406,416,422,436,438,441],[10,11,5],"h1",{"id":12},"精通-comfyui-ltx-23高保真视频生成实用指南",[14,15,16],"p",{},"如果你接触过开源视频生成领域，一定深有体会：串联三十多个节点，祈祷显存不会爆满，等待二十分钟渲染后，得到的视频却出现主体畸形、背景模糊崩坏的问题。",[14,18,19,20,24,25,28],{},"而 LTX-Video 架构的出现，尤其是最新的 LTX 2.3 版本，终于实现了",[21,22,23],"strong",{},"指令遵循度、时序一致性","与",[21,26,27],{},"硬件需求","的平衡。",[14,30,31,32,34],{},"这不是一篇泛泛而谈的概述，我们将在 ComfyUI 中搭建一套稳定实用的**文生视频（T2V）",[21,33,24],{},"图生视频（I2V）**工作流。读完本指南，你将清晰理解每个节点的作用、参数调整逻辑，以及如何避免显卡过载崩溃。",[36,37],"hr",{},[39,40,42],"h2",{"id":41},"一硬件与前置条件真实运行要求","一、硬件与前置条件：真实运行要求",[14,44,45],{},"在开始搭建节点前，先明确本地运行 LTX 2.3 的必备条件。LTX 虽高效，但并非无硬件门槛。",[47,48,49],"h3",{"id":49},"必备配置",[51,52,53,60,66,72],"ul",{},[54,55,56,59],"li",{},[21,57,58],{},"显卡","：推荐 NVIDIA 显卡，显存至少 12GB。8GB 显存可通过极致优化、降低帧数勉强运行，12GB 以上（RTX 3060、4070 及更高型号）可获得流畅体验。",[54,61,62,65],{},[21,63,64],{},"ComfyUI","：更新至最新版本，切勿使用三个月前的旧版本。",[54,67,68,71],{},[21,69,70],{},"ComfyUI Manager","：建议安装该插件，用于快速获取缺失的自定义节点。",[54,73,74,77,78,82],{},[21,75,76],{},"LTX 模型文件","：将 LTX 2.3 核心 safetensors 文件放入 ",[79,80,81],"code",{},"models/checkpoints"," 文件夹，若模型未内置 VAE，需单独下载对应 VAE 文件。",[14,84,85,89],{},[86,87,88],"em",{},"小技巧","：系统虚拟内存至少设置为 32GB。ComfyUI 在权重从内存切换至显存时，过小的虚拟内存会引发看似显存不足的静默崩溃。",[36,91],{},[39,93,95],{"id":94},"二核心原理ltx-23-的独特优势","二、核心原理：LTX 2.3 的独特优势",[14,97,98],{},"如果你用过 AnimateDiff 或 Stable Video Diffusion（SVD），需要转变视频生成思路。",[14,100,101],{},"AnimateDiff 通过滑动上下文窗口强制保证时序一致性，SVD 依靠图像条件预测后续画面，而 LTX 2.3 是基于 DiT（扩散Transformer）架构的原生视频扩散模型。",[14,103,104],{},"对使用者而言，核心差异如下：",[106,107,108,118,124],"ol",{},[54,109,110,113,114,117],{},[21,111,112],{},"提示词逻辑完全不同","：LTX 对动作、镜头运镜、时间节奏的理解远超旧模型，无需堆砌 ",[79,115,116],{},"(masterpiece, best quality, 8k)"," 等词汇，只需像导演一样描述画面。",[54,119,120,123],{},[21,121,122],{},"分辨率严格适配","：Transformer 模型针对特定分辨率与帧数训练，若强行使用 512×512 替代模型适配的 768×512，画面会严重崩坏。",[54,125,126,129],{},[21,127,128],{},"CFG 系数极度敏感","：LTX 的无分类器引导尺度比 SDXL 更苛刻，小幅提升就会导致画面过饱和、噪点过多。",[36,131],{},[39,133,135],{"id":134},"三ltx-23-工作流搭建分步教程","三、LTX 2.3 工作流搭建：分步教程",[14,137,138],{},"从零搭建简洁的文生视频工作流，打开空白 ComfyUI 画布开始操作。",[47,140,142],{"id":141},"步骤1加载基础模型","步骤1：加载基础模型",[14,144,145,146,149,150,153,154,153,157,160],{},"右键画布，选择 ",[79,147,148],{},"添加节点 > 加载器 > Load Checkpoint","，选中 LTX 2.3 模型，该节点会输出 ",[79,151,152],{},"MODEL","、",[79,155,156],{},"CLIP",[79,158,159],{},"VAE","。",[14,162,163,166,167,170],{},[86,164,165],{},"注意","：部分 LTX 适配包提供专属 ",[79,168,169],{},"LTX Model Loader"," 节点，通过 ComfyUI 安装后，建议使用专属加载器，确保 Transformer 模块正常解析。",[47,172,174],{"id":173},"步骤2文本编码器条件设置","步骤2：文本编码器（条件设置）",[14,176,177,178,181,182,160],{},"LTX 依赖高质量文本编码，添加两个 ",[79,179,180],{},"CLIP Text Encode (Prompt)"," 节点，均连接模型输出的 ",[79,183,156],{},[14,185,186,189,190,160],{},[21,187,188],{},"正向提示词（画面与运镜描述）","：摒弃关键词堆砌，使用完整描述语句。\n示例：",[79,191,192],{},"夜间赛博都市霓虹街道上，一辆未来感跑车行驶，电影级广角镜头，摄像机从右向左平移跟随车辆，照片级画质，湿润沥青路面带有反光效果",[14,194,195,198,199,160],{},[21,196,197],{},"反向提示词","：简洁即可。\n",[79,200,201],{},"模糊、变形、畸形、结构错误、低分辨率、静止画面",[47,203,205],{"id":204},"步骤3潜变量配置视频画布设置","步骤3：潜变量配置（视频画布设置）",[14,207,208,209,212],{},"这是最易出错的环节，添加 ",[79,210,211],{},"Empty Latent Video"," 或 LTX 专属潜变量节点。",[14,214,215],{},"LTX 2.3 潜变量黄金规则：",[51,217,218,224,230],{},[54,219,220,223],{},[21,221,222],{},"分辨率","：必须为 32 的倍数，测试首选 768×512 或 512×768，直接使用 1024×576 极易爆显存。",[54,225,226,229],{},[21,227,228],{},"帧数","：建议 17 帧或 33 帧。视频模型需要锚点帧（1+16、1+32），奇数帧数适配性更强。",[54,231,232,235],{},[21,233,234],{},"帧率","：下游可调整，33 帧搭配 8fps 可生成 4 秒流畅短片。",[47,237,239],{"id":238},"步骤4ksampler核心渲染引擎","步骤4：KSampler（核心渲染引擎）",[14,241,242,243,246],{},"添加标准 ",[79,244,245],{},"KSampler"," 节点，连接模型、正负条件、潜变量视频，参数设置如下：",[51,248,249,255,261,271,284,295],{},[54,250,251,254],{},[21,252,253],{},"随机种子","：探索时设为随机，微调成品时固定种子。",[54,256,257,260],{},[21,258,259],{},"迭代步数","：20–30 步，超过 40 步画质无明显提升，仅浪费渲染时间。",[54,262,263,266,267,270],{},[21,264,265],{},"CFG 系数","：",[21,268,269],{},"保持低值","，初始设为 3.0；指令遵循度不足可升至 4.0；画面过曝、出现噪点则降至 2.5。",[54,272,273,266,276,279,280,283],{},[21,274,275],{},"采样器",[79,277,278],{},"euler"," 或 ",[79,281,282],{},"euler_ancestral","，DiT 模型对欧拉采样器适配极佳。",[54,285,286,266,289,279,292,160],{},[21,287,288],{},"调度器",[79,290,291],{},"normal",[79,293,294],{},"sgm_uniform",[54,296,297,300],{},[21,298,299],{},"降噪强度","：纯文生视频设为 1.0。",[47,302,304],{"id":303},"步骤5解码与视频输出","步骤5：解码与视频输出",[14,306,307,308,311,312,315,316,160],{},"将 KSampler 输出的 ",[79,309,310],{},"LATENT"," 连接至 ",[79,313,314],{},"VAE Decode"," 节点，同步连接初始加载的 ",[79,317,159],{},[14,319,320,321,324,325,328],{},"解码后的 ",[79,322,323],{},"IMAGE"," 接入 ",[79,326,327],{},"Video Combine"," 节点（可通过 ComfyUI-VideoHelperSuite 插件获取）：",[51,330,331,334],{},[54,332,333],{},"帧率设为 8、12 或 24，按需选择。",[54,335,336,337,340],{},"格式选择 ",[79,338,339],{},"video/h264-mp4","，支持浏览器原生播放。",[14,342,343,344,347],{},"点击",[21,345,346],{},"队列渲染","，配置正确的情况下，即可生成连贯稳定的短视频。",[36,349],{},[39,351,353],{"id":352},"四进阶技巧图生视频i2v","四、进阶技巧：图生视频（I2V）",[14,355,356],{},"文生视频适合创意探索，图生视频更适合项目实用素材制作。可先在 Midjourney 或 SDXL 生成优质静图，再通过 LTX 赋予动态效果。",[14,358,359],{},"工作流转换步骤：",[106,361,362,369,375,385],{},[54,363,364,365,368],{},"添加 ",[79,366,367],{},"Load Image"," 节点，导入目标静图。",[54,370,364,371,374],{},[79,372,373],{},"VAE Encode"," 节点，编码图像后替换 KSampler 的空潜变量。",[54,376,377,380,381,384],{},[21,378,379],{},"关键设置","：使用 ",[79,382,383],{},"LTX Image Conditioning"," 自定义节点，将图像注入模型或条件流，告知模型以图像为基础生成。",[54,386,387,390],{},[21,388,389],{},"降噪强度调整","：设为 0.8–0.85。设为 1.0 会完全覆盖原图，仅按提示词生成；低于 0.4 则画面无动态效果。",[36,392],{},[39,394,396],{"id":395},"五常见问题排查与修复","五、常见问题排查与修复",[14,398,399],{},"即使配置完美，仍可能出现异常，以下是 LTX 2.3 高频问题解决方案：",[14,401,402,405],{},[21,403,404],{},"问题1：视频中途变为灰色噪点画面","\n解决方法：CFG 系数过高，或提示词包含模型无法理解的动作。将 CFG 降低 0.5，仍异常则简化提示词。",[14,407,408,411,412,415],{},[21,409,410],{},"问题2：主体移动但背景拉伸变形","\n解决方法：典型 DiT 架构瑕疵，添加反向提示词 ",[79,413,414],{},"背景扭曲、透视变形、不合理运镜","，或降低总帧数。模型针对短片段训练，60 帧长视频易超出适配范围。",[14,417,418,421],{},[21,419,420],{},"问题3：CUDA 显存不足（OOM）","\n解决方法：",[106,423,424,427,430,433],{},[54,425,426],{},"关闭占用硬件加速的浏览器标签（如 YouTube）。",[54,428,429],{},"分辨率降至 512×512。",[54,431,432],{},"帧数降至 17 帧。",[54,434,435],{},"通过 ComfyUI-Manager 安装 FP8 量化模型，替代 FP16 原版，显存占用减半，画质几乎无损失。",[36,437],{},[39,439,440],{"id":440},"总结",[14,442,443],{},"在 ComfyUI 中流畅运行 LTX 2.3 需要些许耐心，但回报显著。告别旧模型卡顿、畸形的 AI 视频效果，实现可控、高质量的视频生成。建议从低分辨率、短帧数开始练习，打磨提示词风格，锁定优质种子与构图后，再逐步提升画质与时长。",{"title":445,"searchDepth":446,"depth":446,"links":447},"",2,[448,452,453,460,461,462],{"id":41,"depth":446,"text":42,"children":449},[450],{"id":49,"depth":451,"text":49},3,{"id":94,"depth":446,"text":95},{"id":134,"depth":446,"text":135,"children":454},[455,456,457,458,459],{"id":141,"depth":451,"text":142},{"id":173,"depth":451,"text":174},{"id":204,"depth":451,"text":205},{"id":238,"depth":451,"text":239},{"id":303,"depth":451,"text":304},{"id":352,"depth":446,"text":353},{"id":395,"depth":446,"text":396},{"id":440,"depth":446,"text":440},"md",{},true,"/zh/advanced-tutorial/ltx-workflow",{"title":5,"description":16},"zh/advanced-tutorial/ltx-workflow","bx4e8NXK03H3i15vX6DPO_WmuPBIHfceIuKHfhDq2WU",1773986044745]