<?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[[H3剑心比赛] 5090d 32gb]]></title><description><![CDATA[<h2>视频链接</h2>
<p dir="auto"><a href="https://youtu.be/vV_NDsP8z7I" rel="nofollow ugc">2d</a><br />
<a href="https://youtu.be/sRe_f8WYREQ" rel="nofollow ugc">3d</a></p>
<h2>工作流</h2>
<p dir="auto">在plaugekind的v12基础上，加了三个lora，一个加速lora，两个动作修复lora<br />
<a href="https://upload.lcz.me/uploads/9bab6823-4b36-4b3b-8be0-c7488b7683dc.json" rel="nofollow ugc">PlagueKind-MinimaxH3-V12-fixed.json</a></p>
<h2>生成时间</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>序</th>
<th>开始</th>
<th>结束</th>
<th>ComfyUI 打的 Total Time</th>
<th style="text-align:right">峰值功耗</th>
<th style="text-align:right">峰值显存</th>
</tr>
</thead>
<tbody>
<tr>
<td>1</td>
<td>2026-11-16 11:28:11</td>
<td>2026-11-16 11:33:04</td>
<td><strong>292.81 seconds</strong></td>
<td style="text-align:right">577.59 W</td>
<td style="text-align:right">31766 MiB</td>
</tr>
<tr>
<td>2</td>
<td>2026-11-16 11:28:24</td>
<td>2026-11-16 11:37:48</td>
<td><strong>284.46 seconds</strong></td>
<td style="text-align:right">580.32 W</td>
<td style="text-align:right">31698 MiB</td>
</tr>
</tbody>
</table>
<h2>一、硬件配置</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>项</th>
<th>值</th>
</tr>
</thead>
<tbody>
<tr>
<td>GPU</td>
<td>NVIDIA GeForce RTX 5090 D（单卡 cuda:0）</td>
</tr>
<tr>
<td>显存</td>
<td>32,607 MiB = 31.84 GiB</td>
</tr>
<tr>
<td>驱动</td>
<td>621.14（CUDA 13.0 / cu130）</td>
</tr>
<tr>
<td>功耗上限</td>
<td>575 W 硬件上限 575 W；</td>
</tr>
<tr>
<td>实测峰值功耗</td>
<td>577.59 W（第 1 支）· 580.32 W（第 2 支）</td>
</tr>
<tr>
<td>实测峰值显存</td>
<td>31,766 MiB = 97.4%（第 1 支）· 31,698 MiB（第 2 支）</td>
</tr>
<tr>
<td>实测峰值利用率</td>
<td>99%</td>
</tr>
<tr>
<td>主机内存</td>
<td>100,199,133,184 B ≈ 93.3 GiB</td>
</tr>
<tr>
<td>系统</td>
<td>Windows · ComfyUI portable <code>D:\0.ComfyUI_windows_portable</code></td>
</tr>
<tr>
<td>Python / PyTorch</td>
<td>3.13.14 embedded / 2.14.1+cu130</td>
</tr>
<tr>
<td>ComfyUI</td>
<td>0.45.1（frontend 1.58.5 · comfy-kitchen 0.2.41 · comfy-aimdo 0.6.0）</td>
</tr>
<tr>
<td>启动命令</td>
<td><code>ComfyUI\main.py --windows-standalone-build</code></td>
</tr>
<tr>
<td>显存策略</td>
<td>NORMAL_VRAM · async weight offloading 2 streams · pinned memory 38222</td>
</tr>
</tbody>
</table>
<h2>二、具体工作流</h2>
<p dir="auto">两支是<strong>同一份提交图</strong>（41 个节点），只改了提示词与保存前缀。</p>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>节点 / 参数</th>
<th>值</th>
</tr>
</thead>
<tbody>
<tr>
<td>底模</td>
<td><code>minimax_h3_hybrid_fl2va_ref2va_b25-49-bf16.safetensors</code>（weight_dtype = default）</td>
</tr>
<tr>
<td>文本编码器</td>
<td><code>qwen3vl_32b_minimax_h3_int8_convrot.safetensors</code>（type = minimax）</td>
</tr>
<tr>
<td>视频 VAE</td>
<td><code>lynnreal_omni_light_vae_fp16.safetensors</code></td>
</tr>
<tr>
<td>音频 VAE</td>
<td><code>minimax_h3_audio_vae_fp32.safetensors</code></td>
</tr>
<tr>
<td>采样器</td>
<td>euler · beta_57 · 8 步 · denoise 1.0</td>
</tr>
<tr>
<td>CFG</td>
<td>无 CFG 分支（BasicGuider → SamplerCustomAdvanced）</td>
</tr>
<tr>
<td>SigmaShift</td>
<td>视频 12.0 / 音频 3.0</td>
</tr>
<tr>
<td>分块 FFN</td>
<td>chunks 2 / seq_threshold 4096</td>
</tr>
<tr>
<td>低显存注意力</td>
<td>head_chunks = 4</td>
</tr>
<tr>
<td>参考图</td>
<td>6 张 · ref_image_size = <strong>match</strong></td>
</tr>
<tr>
<td>VLM 编码</td>
<td>vlm_resolution 0 · vlm_video_resolution 256 · reference_fps 24.0</td>
</tr>
<tr>
<td>时长 / fps</td>
<td>12.0 s → 294 帧 · 24.0 fps</td>
</tr>
<tr>
<td>分辨率</td>
<td>参数 1344 × 768 → 实际 <strong>1312 × 736</strong></td>
</tr>
<tr>
<td>保存</td>
<td>SaveVideo → <code>video/ComfyUI</code>，h264 re-encode crf 10</td>
</tr>
</tbody>
</table>
<h3>LoRA 栈</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>LoRA</th>
<th style="text-align:right">强度</th>
</tr>
</thead>
<tbody>
<tr>
<td><code>mmh3\dmad_minimax_h3_8step_full_critic_comfyui.safetensors</code></td>
<td style="text-align:right">1</td>
</tr>
<tr>
<td><code>mmh3\H3_Combat_V2.safetensors</code></td>
<td style="text-align:right">0.7</td>
</tr>
<tr>
<td><code>mmh3\Motion_Repair_V2.safetensors</code></td>
<td style="text-align:right">0.7</td>
</tr>
</tbody>
</table>
<h3>帧数公式（节点 ComfyMathExpression，原句）</h3>
<pre><code>max(5, round(a * 24)) + (5 - (max(5, round(a * 24)) % 17)) % 17
</code></pre>
<p dir="auto"><code>a = 12.0</code> 秒 → <code>round(12*24)=288</code> → 对齐 <code>%17</code> → <strong>294 帧 = 12.25 s @ 24 fps</strong>。</p>
<h3>完整 JSON</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>文件</th>
<th>内容</th>
<th>大小</th>
</tr>
</thead>
<tbody>
<tr>
<td><code>rec/workflow_api_run1.json</code></td>
<td>第 1 支（3D CGI 版）完整 API 图，41 节点</td>
<td>16,747 B</td>
</tr>
<tr>
<td><code>rec/workflow_api_run2.json</code></td>
<td>第 2 支（2D anime 版）完整 API 图，41 节点</td>
<td>16,754 B</td>
</tr>
<tr>
<td><code>rec/history_raw.json</code></td>
<td><code>/history?max_items=50</code> 原始返回（全部节点参数 + 输出元数据）</td>
<td>999,538 B</td>
</tr>
<tr>
<td><code>rec/runs_parsed.json</code></td>
<td>运行结构化摘要（模型/调度/SLA/成片元数据）</td>
<td>8,013 B</td>
</tr>
<tr>
<td><code>rec/gpu-after.csv</code></td>
<td>采集期 nvidia-smi 逐 2 秒采样（118 行）</td>
<td>6,121 B</td>
</tr>
</tbody>
</table>
<p dir="auto">两份 API JSON 的差异只有一处：<code>5479:6012</code> 的 <code>prompt</code>（提示词 Style）。其余 40 个节点逐字节相同。</p>
<h2>三、加速手段</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>手段</th>
<th>状态</th>
<th>证据</th>
</tr>
</thead>
<tbody>
<tr>
<td>CK flash（comfy_kitchen 稠密内核）</td>
<td><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/26a0.png?v=3c59dda5525" class="not-responsive emoji emoji-android emoji--warning" style="height:1.4em;width:auto;vertical-align:middle" title="⚠" alt="⚠" />️ 部分</td>
<td>启动行仍是 <code>Using pytorch attention</code>；SLA 的 <code>dense_backend = comfy_kitchen</code></td>
</tr>
<tr>
<td><strong>SOL Attention（块稀疏 sol_attn）</strong></td>
<td><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/2705.png?v=3c59dda5525" class="not-responsive emoji emoji-android emoji--white_check_mark" style="height:1.4em;width:auto;vertical-align:middle" title="✅" alt="✅" /> 生效</td>
<td><code>engine=comfy_kitchen · sparsity=0.65 · BLK=32x32 · min_seq_len=12288</code></td>
</tr>
<tr>
<td>8 步蒸馏 LoRA（DMAD）</td>
<td><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/2705.png?v=3c59dda5525" class="not-responsive emoji emoji-android emoji--white_check_mark" style="height:1.4em;width:auto;vertical-align:middle" title="✅" alt="✅" /> 生效</td>
<td><code>steps = 8</code>，DMAD str 1.0</td>
</tr>
</tbody>
</table>
]]></description><link>https://lcz.me/topic/3149</link><generator>RSS for Node</generator><lastBuildDate>Thu, 19 Nov 2026 00:16:26 GMT</lastBuildDate><atom:link href="https://lcz.me/topic/3149.rss" rel="self" type="application/rss+xml"/><pubDate>Tue, 17 Nov 2026 09:31:14 GMT</pubDate><ttl>60</ttl><item><title><![CDATA[Reply to [H3剑心比赛] 5090d 32gb on Wed, 18 Nov 2026 00:15:45 GMT]]></title><description><![CDATA[<p dir="auto">你这个生成2d的时候要把参考图中的剑心的3d图片删掉，不然生成的还是3d。提示词压制不过参考图。你试试</p>
]]></description><link>https://lcz.me/post/34232</link><guid isPermaLink="true">https://lcz.me/post/34232</guid><dc:creator><![CDATA[Werther Wang]]></dc:creator><pubDate>Wed, 18 Nov 2026 00:15:45 GMT</pubDate></item><item><title><![CDATA[Reply to [H3剑心比赛] 5090d 32gb on Tue, 17 Nov 2026 11:10:50 GMT]]></title><description><![CDATA[<p dir="auto"><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/1f44f.png?v=3c59dda5525" class="not-responsive emoji emoji-android emoji--clap" style="height:1.4em;width:auto;vertical-align:middle" title=":clap:" alt="👏" /> <img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/1f44f.png?v=3c59dda5525" class="not-responsive emoji emoji-android emoji--clap" style="height:1.4em;width:auto;vertical-align:middle" title=":clap:" alt="👏" /> <img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/1f44f.png?v=3c59dda5525" class="not-responsive emoji emoji-android emoji--clap" style="height:1.4em;width:auto;vertical-align:middle" title=":clap:" alt="👏" /></p>
]]></description><link>https://lcz.me/post/34131</link><guid isPermaLink="true">https://lcz.me/post/34131</guid><dc:creator><![CDATA[xiongjie gao]]></dc:creator><pubDate>Tue, 17 Nov 2026 11:10:50 GMT</pubDate></item><item><title><![CDATA[Reply to [H3剑心比赛] 5090d 32gb on Tue, 17 Nov 2026 10:51:18 GMT]]></title><description><![CDATA[<p dir="auto"><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/2705.png?v=3c59dda5525" class="not-responsive emoji emoji-android emoji--white_check_mark" style="height:1.4em;width:auto;vertical-align:middle" title="✅" alt="✅" /> 版主审定：本帖设为<strong>优秀</strong>，作者 <strong>+3 积分</strong>奖励，继续加油！</p>
<blockquote>
<p dir="auto">奖励凭证：优秀 +3 分 · 编号 e3149-3611-1</p>
</blockquote>
]]></description><link>https://lcz.me/post/34125</link><guid isPermaLink="true">https://lcz.me/post/34125</guid><dc:creator><![CDATA[Xiaote]]></dc:creator><pubDate>Tue, 17 Nov 2026 10:51:18 GMT</pubDate></item><item><title><![CDATA[Reply to [H3剑心比赛] 5090d 32gb on Tue, 17 Nov 2026 10:51:15 GMT]]></title><description><![CDATA[<p dir="auto">相当牛逼，过审。<br />
20积分比赛结束后统一结算。</p>
]]></description><link>https://lcz.me/post/34124</link><guid isPermaLink="true">https://lcz.me/post/34124</guid><dc:creator><![CDATA[terry]]></dc:creator><pubDate>Tue, 17 Nov 2026 10:51:15 GMT</pubDate></item></channel></rss>