<?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] RTX PRO5000 MiniMax H3 2D / 3D]]></title><description><![CDATA[<h2>[H3剑心比赛] RTX PRO5000 MiniMax H3 2D / 3D</h2>
<h2>先说结论</h2>
<ul>
<li>
<p dir="auto"><strong>12 秒 / 294 帧 / 24fps，2D 与 3D 各一条，一次成型</strong>，走同一套多参考图工作流。</p>
</li>
<li>
<p dir="auto">两条<strong>提示词逐字相同</strong>（只有第一行的风格词 <code>2D traditional anime</code> / <code>3D CGI animation</code> 不同）、<strong>seed 相同</strong>（<code>813788674966030</code>）、<strong>步数相同（12）</strong>，唯一变量就是风格——所以这是一组干净的对照。</p>
</li>
<li>
<p dir="auto">实测耗时：<strong>2D 678 秒（11:18）</strong>，<strong>3D 721 秒（12:00）</strong>，均为整段一次跑完（含加载与解码）。</p>
</li>
<li>
<p dir="auto">加速不靠偷步数，靠<strong>蒸馏 LoRA + SageAttention + INT8 注意力后端</strong>三层。</p>
</li>
<li>
<p dir="auto">YouTube：<br />
H3剑心  2D   <a href="https://youtu.be/r8NvaIm_0NU" rel="nofollow ugc">https://youtu.be/r8NvaIm_0NU</a><br />
H3剑心  3D   <a href="https://youtu.be/DDn4dBp-0W4" rel="nofollow ugc">https://youtu.be/DDn4dBp-0W4</a></p>
</li>
</ul>
<h2>1. 硬件与软件</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 RTX PRO 5000 Blackwell，48GB（<code>nvidia-smi</code> 实测 48,935 MiB）</td>
</tr>
<tr>
<td>驱动</td>
<td>621.14</td>
</tr>
<tr>
<td>显存占用</td>
<td>单条 12 秒任务常驻约 34GB</td>
</tr>
<tr>
<td>系统内存</td>
<td>64GB</td>
</tr>
<tr>
<td>Python / Torch</td>
<td>3.13.12 / <strong>torch 2.12.1+cu130</strong></td>
</tr>
<tr>
<td>ComfyUI</td>
<td>0.29（桌面版）</td>
</tr>
<tr>
<td>加速运行时</td>
<td>SageAttention 2.2.0+cu130、triton-windows 3.9.0</td>
</tr>
</tbody>
</table>
<h2>2. 工作流</h2>
<ul>
<li>文件：**<a href="https://upload.lcz.me/uploads/231a7fed-9e59-4870-ae8f-fa5a5f0ada8c.json" rel="nofollow ugc">MiniMaxH3多参武戏生成工作流_秒数版_直填分辨率V2.json</a></li>
<li>主链：<code>UNETLoader → LoraLoaderModelOnly（加速）→ LoraLoaderModelOnly（武戏）→ MiniMaxH3MemoryEfficientSageAttentionPatch → 引导器 → BasicScheduler(beta) → SamplerCustomAdvanced → VAEDecode → VHS_VideoCombine</code></li>
<li>生成节点：<code>MiniMaxH3ReferenceToVideo</code>（Ref2VA，多参考图 + 模型自生成音轨）</li>
<li><strong>帧数不用手算</strong>：<code>秒数(12.0) → ComfyMathExpression → 294 帧</code>，公式 <code>max(5, round(秒×24))</code>，再补到 17k+5</li>
<li>参考图：6张</li>
</ul>
<h2>3. 加速手段</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>手段</th>
<th>位置</th>
<th>作用</th>
</tr>
</thead>
<tbody>
<tr>
<td><strong>turbo 8 步蒸馏 LoRA</strong> <code>minimax_h3_ref2v_turbo_8step_v1.0_768p</code> @ <strong>1.0</strong></td>
<td>链首</td>
<td>把官方 25 步压到 8–12 步，提速主力</td>
</tr>
<tr>
<td><strong>SageAttention 补丁</strong> <code>MiniMaxH3MemoryEfficientSageAttentionPatch</code></td>
<td>每个 H3 block</td>
<td>int8 注意力内核，日志确认逐 block 生效</td>
</tr>
<tr>
<td><strong>INT8 注意力后端</strong> <code>ModelAttentionBackend = comfy kitchen attention</code></td>
<td>采样前</td>
<td>ComfyUI 自带量化注意力</td>
</tr>
<tr>
<td>低显存分块 <code>LowVRAMAttention(head_chunks 2)</code> + <code>ChunkFeedForward(2, 4096)</code></td>
<td>采样前</td>
<td>不加速，只为 48G 上稳跑不 OOM</td>
</tr>
</tbody>
</table>
<ul>
<li>未用 SOL Attention（本机未装）与 FlashAttention 补丁（未启用）。</li>
<li>蒸馏 LoRA 配 8–12 步都在轨；本机对比过 8 / 10 / 12 步，<strong>细节随步数继续涨、耗时增量很小</strong>，故终版统一 12 步。</li>
</ul>
<h2>4. 完整运行日志</h2>
<p dir="auto"><img src="https://upload.lcz.me/uploads/32d7c5a8-5d03-4098-a7dc-b789fdcde950.jpg" alt="img_v3_0217e_1c1170d2-440f-4f89-8386-30b693d5567g.jpg" class="img-markdown  img-fluid" /></p>
<p dir="auto">**<br />
<img src="https://upload.lcz.me/uploads/d3307e07-359a-476c-aaeb-54a9752a1403.jpg" alt="img_v3_0217e_8cbef899-d1ef-4e5d-815a-7e08fdd1b84g.jpg" class="img-markdown  img-fluid" /><br />
控制台截图**。</p>
<h2>5. 规格证明</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>项</th>
<th>2D</th>
<th>3D</th>
</tr>
</thead>
<tbody>
<tr>
<td>工作流</td>
<td>秒数版（直填分辨率 V2）</td>
<td>秒数版（直填分辨率 V2）</td>
</tr>
<tr>
<td>步数</td>
<td><strong>12</strong></td>
<td><strong>12</strong></td>
</tr>
<tr>
<td>采样器</td>
<td>beta / euler / denoise 1.0</td>
<td>同左</td>
</tr>
<tr>
<td>LoRA</td>
<td>turbo_8step @1.0 + <code>H3_Fight_Flow_V1</code> @0.6</td>
<td>同左</td>
</tr>
<tr>
<td>分辨率</td>
<td><strong>1344×768</strong></td>
<td><strong>1344×768</strong></td>
</tr>
<tr>
<td>帧数</td>
<td>294</td>
<td>294</td>
</tr>
<tr>
<td>时长</td>
<td>12.250s @24fps</td>
<td>12.250s @24fps</td>
</tr>
<tr>
<td>音轨</td>
<td>aac 2ch 12.224s（模型自生成）</td>
<td>同左</td>
</tr>
<tr>
<td>seed</td>
<td>813788674966030</td>
<td>813788674966030</td>
</tr>
<tr>
<td>参考图</td>
<td>5 张</td>
<td>6 张</td>
</tr>
<tr>
<td>采样耗时</td>
<td>10:18（末步 51.58 s/it）</td>
<td>11:01（末步 55.12 s/it）</td>
</tr>
<tr>
<td>总耗时</td>
<td>678 秒（11:18）</td>
<td>721 秒（12:00）</td>
</tr>
<tr>
<td>文件体积</td>
<td>12.5 MB</td>
<td>10.2 MB</td>
</tr>
</tbody>
</table>
<p dir="auto"><code>ffprobe</code> 实测（两条输出一致）：</p>
<pre><code>$ ffprobe -v error -select_streams v:0 -show_entries stream=width,height,nb_frames,r_frame_rate,duration \
    -of default=nw=1 视频/H3武戏_2D_12步_00009.mp4
width=1344
height=768
r_frame_rate=24/1
duration=12.250000
nb_frames=294

$ ffprobe -v error -select_streams a:0 -show_entries stream=codec_name,channels,duration \
    -of default=nw=1 视频/H3武戏_2D_12步_00009.mp4
codec_name=aac
channels=2
duration=12.224000
</code></pre>
<h2>6. 复现清单</h2>
<ol>
<li>底模 <code>minimax_h3_ref2va_pruned_fp8_scaled</code>（FLOW_AV / bf16）</li>
<li>加速 LoRA <code>minimax_h3_ref2v_turbo_8step_v1.0_768p_comfyui_bf16</code> @1.0</li>
<li>武戏 LoRA <code>H3_Fight_Flow_V1</code> @0.6</li>
<li><code>BasicScheduler</code> = beta，steps 12，denoise 1.0，采样器 euler</li>
<li><code>194 秒数</code> 填 12 → 294 帧；<code>196 分辨率</code> 填 1344×768、倍数 32</li>
<li>参考图按 <code>ref_image_0</code>～<code>ref_image_5</code> 顺序喂入（3D 6 张 / 2D 5 张）</li>
<li>注意力：Sage 补丁 + kitchen attention；显存不够再开 LowVRAMAttention / ChunkFeedForward</li>
</ol>
]]></description><link>https://lcz.me/topic/3069</link><generator>RSS for Node</generator><lastBuildDate>Thu, 19 Nov 2026 00:36:40 GMT</lastBuildDate><atom:link href="https://lcz.me/topic/3069.rss" rel="self" type="application/rss+xml"/><pubDate>Sat, 14 Nov 2026 14:46:06 GMT</pubDate><ttl>60</ttl><item><title><![CDATA[Reply to [H3剑心比赛] RTX PRO5000 MiniMax H3 2D / 3D on Sat, 14 Nov 2026 17:11:54 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 e3069-5250-1</p>
</blockquote>
]]></description><link>https://lcz.me/post/33354</link><guid isPermaLink="true">https://lcz.me/post/33354</guid><dc:creator><![CDATA[Xiaote]]></dc:creator><pubDate>Sat, 14 Nov 2026 17:11:54 GMT</pubDate></item><item><title><![CDATA[Reply to [H3剑心比赛] RTX PRO5000 MiniMax H3 2D / 3D on Sat, 14 Nov 2026 17:11:50 GMT]]></title><description><![CDATA[<p dir="auto">通过初审，正式参赛，奖励20积分，比赛结束后统一结算！</p>
]]></description><link>https://lcz.me/post/33353</link><guid isPermaLink="true">https://lcz.me/post/33353</guid><dc:creator><![CDATA[terry]]></dc:creator><pubDate>Sat, 14 Nov 2026 17:11:50 GMT</pubDate></item></channel></rss>