<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title><![CDATA[ubuntu26.04下7900xtx跑comfyui工作流阶段总结]]></title><description><![CDATA[<p dir="auto">走了些弯路，折腾了一段时间wan模型，听锤哥的果断换了ltx，目前flux和ltx2.3还有数字人都跑通了，数字人是下载刘悦的在本地修改后跑通的，效果还不错。</p>
<h1>ComfyUI 工作流成果总结</h1>
<blockquote>
<p dir="auto">最后更新：2026-05-18</p>
</blockquote>
<hr />
<h2>硬件环境 Hardware</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>组件</th>
<th>规格</th>
</tr>
</thead>
<tbody>
<tr>
<td>CPU</td>
<td>Intel Core i5-12400F (12核)</td>
</tr>
<tr>
<td>内存</td>
<td>48GB (MemoryMax,)</td>
</tr>
<tr>
<td>GPU</td>
<td>AMD Radeon RX 7900 XTX (24GB VRAM)</td>
</tr>
<tr>
<td>系统盘</td>
<td>Kingston NV2 1TB NVMe (279G分区, 39G余量)</td>
</tr>
<tr>
<td>外置盘</td>
<td>zyz盘 448GB exfat (94%满, 已挂载)</td>
</tr>
<tr>
<td>ROCm</td>
<td>ROCm 7.2</td>
</tr>
<tr>
<td>PyTorch</td>
<td>2.12.0+rocm7.2</td>
</tr>
</tbody>
</table>
<h2>软件环境 Software</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>组件</th>
<th>版本/配置</th>
</tr>
</thead>
<tbody>
<tr>
<td>OS</td>
<td>Linux 7.0.0-15-generic</td>
</tr>
<tr>
<td>ComfyUI</td>
<td>v0.20.1+ (API + 浏览器)</td>
</tr>
<tr>
<td>ComfyUI 端口</td>
<td>8188 (--listen 0.0.0.0)</td>
</tr>
<tr>
<td>启动参数</td>
<td><code>--disable-async-offload</code> (不加--lowvram)</td>
</tr>
<tr>
<td>环境变量</td>
<td><code>TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=0</code></td>
</tr>
</tbody>
</table>
<hr />
<h2>工作流总览</h2>
<p dir="auto">共 5 个工作流，全部存在 <code>/root/ComfyUI/workflows_hermes/</code></p>
<h3>1. Flux1.dev-fp8 文生图</h3>
<p dir="auto"><strong>文件</strong>: <code>flux1_t2i.json</code></p>
<p dir="auto"><strong>模型与位置</strong>:</p>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>模型</th>
<th>类型</th>
<th>大小</th>
<th>位置</th>
</tr>
</thead>
<tbody>
<tr>
<td>flux1-dev-fp8.safetensors</td>
<td>checkpoint (模型+CLIP+VAE)</td>
<td>17GB</td>
<td><code>/root/ComfyUI/models/checkpoints/</code></td>
</tr>
</tbody>
</table>
<p dir="auto"><strong>参数设置</strong>:</p>
<ul>
<li>分辨率: 1024×1024</li>
<li>采样器: euler</li>
<li>调度器: normal</li>
<li>步数: 20</li>
<li>CFG: 3.5</li>
<li>去噪: 1.0</li>
</ul>
<p dir="auto"><strong>节点结构</strong>: CheckpointLoaderSimple → CLIPTextEncode×2 + EmptyLatentImage → KSampler → VAEDecode → SaveImage</p>
<hr />
<h3>2. Flux2 Klein 4B 文生图</h3>
<p dir="auto"><strong>文件</strong>: <code>flux2_t2i.json</code></p>
<p dir="auto"><strong>模型与位置</strong>:</p>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>模型</th>
<th>类型</th>
<th>大小</th>
<th>位置</th>
</tr>
</thead>
<tbody>
<tr>
<td>flux-2-klein-4b.safetensors</td>
<td>U-Net (diffusion model)</td>
<td>7.3GB</td>
<td><code>/root/ComfyUI/models/diffusion_models/</code></td>
</tr>
<tr>
<td>qwen_3_4b.safetensors</td>
<td>CLIP/T5文本编码器</td>
<td>7.5GB</td>
<td><code>/root/ComfyUI/models/text_encoders/</code></td>
</tr>
<tr>
<td>flux2-vae.safetensors</td>
<td>VAE解码器</td>
<td>321MB</td>
<td><code>/root/ComfyUI/models/vae/</code></td>
</tr>
</tbody>
</table>
<p dir="auto"><strong>参数设置</strong>:</p>
<ul>
<li>分辨率: 1024×1024</li>
<li>采样器: euler</li>
<li>调度器: simple</li>
<li>步数: 20</li>
<li>CFG: 3.5</li>
<li>去噪: 1.0</li>
</ul>
<p dir="auto"><strong>节点结构</strong>: UNETLoader + CLIPLoader(type=flux2) + VAELoader → CLIPTextEncode×2 + EmptyLatentImage → KSampler → VAEDecode → SaveImage</p>
<blockquote>
<p dir="auto"><strong>注意</strong>: 文本编码器必须用 <code>qwen_3_4b.safetensors</code> + CLIPLoader(type="flux2")。不要用DualCLIPLoader或UMT5，那会报错。</p>
</blockquote>
<hr />
<h3>3. LTX 2.3 文生视频 Text-to-Video</h3>
<p dir="auto"><strong>文件</strong>: <code>ltx23_t2v.json</code></p>
<p dir="auto"><strong>模型与位置</strong>:</p>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>模型</th>
<th>类型</th>
<th>大小</th>
<th>位置</th>
</tr>
</thead>
<tbody>
<tr>
<td>ltx-2.3-22b-distilled-1.1-Q3_K_M.gguf</td>
<td>U-Net GGUF</td>
<td>9.9GB</td>
<td><code>/root/ComfyUI/models/unet/</code> 本地</td>
</tr>
<tr>
<td>gemma-3-12b-it-Q4_K_M.gguf</td>
<td>CLIP文本编码器 GGUF</td>
<td>6.8GB</td>
<td><code>/root/ComfyUI/models/clip/</code> 本地</td>
</tr>
<tr>
<td>ltx-2.3_text_projection_bf16.safetensors</td>
<td>文本投影</td>
<td>75B</td>
<td><code>/root/ComfyUI/models/clip/</code> (symlink)</td>
</tr>
<tr>
<td>LTX23_video_vae_bf16.safetensors</td>
<td>VAE</td>
<td>1.4GB</td>
<td><code>/root/ComfyUI/models/vae/</code></td>
</tr>
</tbody>
</table>
<p dir="auto"><strong>参数设置</strong>:</p>
<ul>
<li>分辨率: 544×960</li>
<li>帧数: 49 (约2秒)</li>
<li>帧率: 24fps</li>
<li>采样器: euler</li>
<li>步数: 8</li>
<li>CFG: 1.0</li>
<li>max_shift: 2.05, base_shift: 0.95</li>
<li>VAE分块: 4×4 tiles, overlap=2</li>
<li>输出: MP4 H.264</li>
</ul>
<p dir="auto"><strong>节点结构</strong>: UnetLoaderGGUF + DualCLIPLoaderGGUF(type=ltxv) + VAELoader → EmptyLTXVLatentVideo → CLIPTextEncode×2 → LTXVConditioning → LTXVScheduler → CFGGuider + RandomNoise + KSamplerSelect → SamplerCustomAdvanced → LTXVCropGuides → LTXVTiledVAEDecode → VHS_VideoCombine</p>
<hr />
<h3>4. LTX 2.3 图生视频 Image-to-Video</h3>
<p dir="auto"><strong>文件</strong>: <code>ltx23_i2v.json</code></p>
<p dir="auto"><strong>模型与位置</strong> (同T2V基础上增加):</p>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>模型</th>
<th>类型</th>
<th>大小</th>
<th>位置</th>
</tr>
</thead>
<tbody>
<tr>
<td>同上 +</td>
<td></td>
<td></td>
<td></td>
</tr>
<tr>
<td>ltx-2-19b-lora-camera-control-static.safetensors</td>
<td>LoRA</td>
<td>2.1GB</td>
<td><code>/root/ComfyUI/models/loras/</code> 本地</td>
</tr>
</tbody>
</table>
<p dir="auto"><strong>参数设置</strong>:</p>
<ul>
<li>分辨率: 544×960</li>
<li>帧数: 49 (约2秒)</li>
<li>帧率: 24fps</li>
<li>采样器: euler, 8步</li>
<li>CFG: 1.0</li>
<li>静态相机LoRA: strength=0.8</li>
<li>Inplace注入: strength=1.0</li>
<li>输入: <code>your_photo.png</code> (自动resize到544×960)</li>
</ul>
<p dir="auto"><strong>节点结构</strong>: UnetLoaderGGUF + DualCLIPLoaderGGUF → LoraLoaderModelOnly + LoadImage → ImageResizeKJv2 + EmptyLTXVLatentVideo → LTXVImgToVideoInplace → (同T2V采样→解码→合成)</p>
<hr />
<h3>5. LTX 2.3 数字人 LipSync (图+音频→视频)</h3>
<p dir="auto"><strong>文件</strong>: <code>ltx23_lipsync_digitalhuman.json</code></p>
<p dir="auto"><strong>模型与位置</strong> (同I2V基础上增加):</p>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>模型</th>
<th>类型</th>
<th>大小</th>
<th>位置</th>
</tr>
</thead>
<tbody>
<tr>
<td>ltx23_q4ks.gguf (LTX-2.3-distilled-Q4_K_S.gguf)</td>
<td>U-Net GGUF</td>
<td>16GB</td>
<td><code>/root/ComfyUI/models/unet/</code> 软链→zyz盘</td>
</tr>
<tr>
<td>MelBandRoformer_fp16.safetensors</td>
<td>音频人声分离</td>
<td>436MB</td>
<td><code>/root/ComfyUI/models/diffusion_models/</code> 软链→MelBandRoformer/</td>
</tr>
<tr>
<td>其他模型同I2V</td>
<td></td>
<td></td>
<td></td>
</tr>
</tbody>
</table>
<p dir="auto"><strong>参数设置</strong>:</p>
<ul>
<li>分辨率: 544×960</li>
<li>帧数: 73 (约3秒)</li>
<li>帧率: 24fps</li>
<li>采样器: euler, <strong>12步</strong></li>
<li>CFG: 1.0</li>
<li>静态相机LoRA: strength=0.8</li>
<li>输入: <code>your_photo.png</code> + <code>your_audio.wav</code></li>
<li>音频处理: MelBandRoFormer 分离人声 → VHS_VideoCombine.audio 直接合并</li>
<li>音频时长裁剪: 3.0秒 (匹配73帧@24fps)</li>
</ul>
<blockquote>
<p dir="auto"><strong>重要</strong>: 音频不走latent注入（LTXVAudioVAEEncode输出NestedTensor不兼容GGUF），而是走MelBandRoFormer分离后直拼。VHS_VideoCombine的<code>pingpong</code>为必填参数，缺了会导致静默失败。</p>
</blockquote>
<hr />
<h2>VRAM占用估算</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>工作流</th>
<th>模型</th>
<th>总VRAM</th>
<th>可用</th>
</tr>
</thead>
<tbody>
<tr>
<td>Flux1.dev-fp8</td>
<td>17GB checkpoint</td>
<td>~17GB</td>
<td><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/2705.png?v=d348ca29232" class="not-responsive emoji emoji-android emoji--white_check_mark" style="height:23px;width:auto;vertical-align:middle" title="✅" alt="✅" /> 余~7GB</td>
</tr>
<tr>
<td>Flux2 Klein 4B</td>
<td>7.3G+7.5G+321M</td>
<td>~15GB</td>
<td><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/2705.png?v=d348ca29232" class="not-responsive emoji emoji-android emoji--white_check_mark" style="height:23px;width:auto;vertical-align:middle" title="✅" alt="✅" /> 余~9GB</td>
</tr>
<tr>
<td>LTX T2V (Q3_K_M)</td>
<td>9.9G+6.8G</td>
<td>~17GB</td>
<td><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/2705.png?v=d348ca29232" class="not-responsive emoji emoji-android emoji--white_check_mark" style="height:23px;width:auto;vertical-align:middle" title="✅" alt="✅" /> 余~7GB</td>
</tr>
<tr>
<td>LTX I2V (Q3_K_M)</td>
<td>9.9G+6.8G+2.1G</td>
<td>~19GB</td>
<td><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/2705.png?v=d348ca29232" class="not-responsive emoji emoji-android emoji--white_check_mark" style="height:23px;width:auto;vertical-align:middle" title="✅" alt="✅" /> 余~5GB</td>
</tr>
<tr>
<td>LTX 数字人 (Q4_K_S)</td>
<td>16G+6.8G+2.1G+436M</td>
<td>~23GB</td>
<td><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/26a0.png?v=d348ca29232" class="not-responsive emoji emoji-android emoji--warning" style="height:23px;width:auto;vertical-align:middle" title="⚠" alt="⚠" />️ 余~1GB</td>
</tr>
</tbody>
</table>
<blockquote>
<p dir="auto">Q4_K_S 数字人峰值 ~23GB 非常紧，建议先用 Q3_K_M 测试再换 Q4_K_S。</p>
</blockquote>
<h2>最大视频时长 (480p 720×480)</h2>
<p dir="auto">将latent从544×960改为720×480，latent缩小~35%：</p>
<ul>
<li>Q3_K_M: ~12秒 (290帧)</li>
<li>Q4_K_S: ~8秒 (193帧)</li>
</ul>
<hr />
<h2>已知问题</h2>
<ol>
<li><strong>动作迁移 (ICLoRA)</strong>: 需要原生FP8模型22GB，24GB VRAM装不下。ICLoRA不兼容GGUF路径，当前无法实现完整的动作迁移。</li>
<li><strong>音频latent注入</strong>: LTXVAudioVAEEncode输出NestedTensor，与GGUF U-Net不兼容。已绕过走MeldBandRoFormer直拼。</li>
<li><strong>VHS_VideoCombine</strong>: <code>pingpong</code>必填，缺了静默失败。<code>LTXVSeparateAVLatent</code>无音频时访问<code>latents[1]</code>会IndexError——I2V应跳过此节点。</li>
<li><strong>Flux2文本编码器</strong>: 必须用 <code>qwen_3_4b.safetensors</code> + CLIPLoader(type="flux2")。不能用DualCLIPLoader。</li>
<li><strong>zyz盘依赖</strong>: 数字人工作流的Q4_K_S模型在zyz盘软链，拔盘后工作流不可用。Q3_K_M版在本地可独立运行。</li>
</ol>
<hr />
]]></description><link>https://lcz.me/topic/196/ubuntu26.04下7900xtx跑comfyui工作流阶段总结</link><generator>RSS for Node</generator><lastBuildDate>Wed, 20 May 2026 06:58:40 GMT</lastBuildDate><atom:link href="https://lcz.me/topic/196.rss" rel="self" type="application/rss+xml"/><pubDate>Mon, 18 May 2026 07:27:57 GMT</pubDate><ttl>60</ttl><item><title><![CDATA[Reply to ubuntu26.04下7900xtx跑comfyui工作流阶段总结 on Wed, 20 May 2026 03:40:59 GMT]]></title><description><![CDATA[<p dir="auto">谢谢作者分享，我也回去继续折腾我的7900XTX</p>
]]></description><link>https://lcz.me/post/2711</link><guid isPermaLink="true">https://lcz.me/post/2711</guid><dc:creator><![CDATA[陈鸿]]></dc:creator><pubDate>Wed, 20 May 2026 03:40:59 GMT</pubDate></item><item><title><![CDATA[Reply to ubuntu26.04下7900xtx跑comfyui工作流阶段总结 on Tue, 19 May 2026 08:35:22 GMT]]></title><description><![CDATA[<p dir="auto">感謝大大的分享~努力學習中</p>
]]></description><link>https://lcz.me/post/2563</link><guid isPermaLink="true">https://lcz.me/post/2563</guid><dc:creator><![CDATA[sil041]]></dc:creator><pubDate>Tue, 19 May 2026 08:35:22 GMT</pubDate></item><item><title><![CDATA[Reply to ubuntu26.04下7900xtx跑comfyui工作流阶段总结 on Tue, 19 May 2026 08:18:21 GMT]]></title><description><![CDATA[<p dir="auto">看来必须要用ubuntu</p>
]]></description><link>https://lcz.me/post/2554</link><guid isPermaLink="true">https://lcz.me/post/2554</guid><dc:creator><![CDATA[janebo]]></dc:creator><pubDate>Tue, 19 May 2026 08:18:21 GMT</pubDate></item><item><title><![CDATA[Reply to ubuntu26.04下7900xtx跑comfyui工作流阶段总结 on Mon, 18 May 2026 08:26:51 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/%E5%BC%A0%E8%80%81%E5%B8%88" aria-label="Profile: 张老师">@<bdi>张老师</bdi></a> <img src="https://upload.lcz.me/uploads/5bd08928-c7bb-4cb5-8956-92bba313050b.jpeg" alt="9041754d-ef62-47e4-9419-051606069f96-image.jpeg" class=" img-fluid img-markdown" /></p>
]]></description><link>https://lcz.me/post/2333</link><guid isPermaLink="true">https://lcz.me/post/2333</guid><dc:creator><![CDATA[terry]]></dc:creator><pubDate>Mon, 18 May 2026 08:26:51 GMT</pubDate></item><item><title><![CDATA[Reply to ubuntu26.04下7900xtx跑comfyui工作流阶段总结 on Mon, 18 May 2026 08:07:11 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/terry" aria-label="Profile: terry">@<bdi>terry</bdi></a> 我发现论坛没有收藏功能，好几个帖子，我回头想来具体看看，但是找半天找不到，以后帖子越来越多……</p>
]]></description><link>https://lcz.me/post/2321</link><guid isPermaLink="true">https://lcz.me/post/2321</guid><dc:creator><![CDATA[张老师]]></dc:creator><pubDate>Mon, 18 May 2026 08:07:11 GMT</pubDate></item><item><title><![CDATA[Reply to ubuntu26.04下7900xtx跑comfyui工作流阶段总结 on Mon, 18 May 2026 07:57:10 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/weidong" aria-label="Profile: weidong">@<bdi>weidong</bdi></a> 大哥，正常就是这样的。</p>
]]></description><link>https://lcz.me/post/2316</link><guid isPermaLink="true">https://lcz.me/post/2316</guid><dc:creator><![CDATA[terry]]></dc:creator><pubDate>Mon, 18 May 2026 07:57:10 GMT</pubDate></item><item><title><![CDATA[Reply to ubuntu26.04下7900xtx跑comfyui工作流阶段总结 on Mon, 18 May 2026 07:56:20 GMT]]></title><description><![CDATA[<p dir="auto">我这工作流很奇怪，那些连线总是显示不出来，但点一下那些框框就出来了，不影响跑<br />
<img src="https://upload.lcz.me/uploads/5cedb34d-5b85-4a08-a298-e66ad382193c.png" alt="微信图片_20260518155445_510_749.png" class=" img-fluid img-markdown" /></p>
]]></description><link>https://lcz.me/post/2315</link><guid isPermaLink="true">https://lcz.me/post/2315</guid><dc:creator><![CDATA[weidong]]></dc:creator><pubDate>Mon, 18 May 2026 07:56:20 GMT</pubDate></item><item><title><![CDATA[Reply to ubuntu26.04下7900xtx跑comfyui工作流阶段总结 on Mon, 18 May 2026 07:31:37 GMT]]></title><description><![CDATA[<p dir="auto">很详细，精品，补点截图，我给置顶。</p>
]]></description><link>https://lcz.me/post/2305</link><guid isPermaLink="true">https://lcz.me/post/2305</guid><dc:creator><![CDATA[terry]]></dc:creator><pubDate>Mon, 18 May 2026 07:31:37 GMT</pubDate></item></channel></rss>