<?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[MiniMax H3 视频大赛参赛帖：双 R9700 本地管线 1344×768/12s 全质量出片]]></title><description><![CDATA[<h2>0. 结论</h2>
<ul>
<li>双卡 R9700（32G/卡）本地跑通 H3 R2V 全链路：<strong>6 张官方参考图 → 1344×768 / 12.25s / 294 帧 / 20 步全质量</strong>，单支 <strong>76.5 分钟</strong>，3D/2D 两版同提示词同 seed。</li>
<li>加速构成 = <strong>int8 convrot 量化模型（20G 主模型）+ comfy_kitchen HIP 内核</strong>；实测三种注意力机制无差异（见 §5），瓶颈在量化 Linear 与模型换装，不在 attention。</li>
<li>关键坑一个：H3 的 <code>ref_images</code>（AUTOGROW）在 /prompt API 必须用<strong>扁平点号键</strong>，常规 list 写法被静默丢弃 → 退化为纯文生视频、参考图完全失效。解法见 §4。</li>
</ul>
<h2>1. 硬件与软件</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>项</th>
<th>配置</th>
</tr>
</thead>
<tbody>
<tr>
<td>主板/CPU</td>
<td>X99-TF / E5-2696 v3（18C36T）</td>
</tr>
<tr>
<td>内存</td>
<td>128G DDR4 ECC</td>
</tr>
<tr>
<td>GPU</td>
<td>2× AMD Radeon R9700 32G（gfx1201）</td>
</tr>
<tr>
<td>系统</td>
<td>Ubuntu 26.04.1 / kernel 7.0.0-43 / ROCm 7.2.4</td>
</tr>
<tr>
<td>ComfyUI</td>
<td>前端 1.58.5，comfy_kitchen hip 后端</td>
</tr>
<tr>
<td>电源</td>
<td>1250W（双卡直连）</td>
</tr>
</tbody>
</table>
<h2>2. 模型清单</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>文件</th>
<th>用途</th>
<th>体积</th>
</tr>
</thead>
<tbody>
<tr>
<td>minimax_h3_ref2va_pruned_int8_convrot.safetensors</td>
<td>R2V 主干（DiT）</td>
<td>20G</td>
</tr>
<tr>
<td>qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors</td>
<td>文本/参考图编码器（VLM）</td>
<td>15.7G</td>
</tr>
<tr>
<td>minimax_h3_video_vae_int8_convrot.safetensors</td>
<td>视频 VAE</td>
<td>—</td>
</tr>
<tr>
<td>minimax_h3_audio_vae_fp32.safetensors</td>
<td>音频 VAE（原生音轨）</td>
<td>—</td>
</tr>
</tbody>
</table>
<p dir="auto">单卡显存账：TE 15.7G + 主模型 20G = 35.7G &gt; 32G，ComfyUI 自动 offload 换装（<code>loaded partially, lowvram patches</code> 属正常）。</p>
<h2>3. 工作流（/prompt API 提交，可复现）</h2>
<pre><code>LoadImage×6 → MiniMaxH3ReferenceToVideo(ref_image_size=match)
  ├─ clip  = CLIPLoader(qwen3vl_32b nvfp4, type=minimax)
  ├─ vae   = video VAE / audio_vae = audio VAE
  ├─ width=1344 height=768 length=288
  └─ 输出: conditioning + 空白视频 latent
RandomNoise(seed=12345) → KSamplerSelect(res_multistep)
BasicScheduler(model=主干, simple, steps=20, denoise=1.0)   ← 无 LoRA 全质量档
BasicGuider → SamplerCustomAdvanced
VAEDecode + VAEDecodeAudio → CreateVideo(24fps) → SaveVideo
</code></pre>
<p dir="auto">提示词（3D/2D 两版仅风格行不同）：</p>
<pre><code>(Style: 3D CGI animation)      ← 2D 版为 (Style: 2D traditional anime)
Masterpiece, high quality, cinematic lighting.
Subject: A fierce battle sequence featuring two characters based on the
reference images—a red-clad swordsman (Jianxin) and a purple-robed warrior
(Zhizhixiong).
Action: The camera begins with a dynamic low-angle tracking shot. The swordsman
in red dashes forward with a long sword, leaving a glowing trail, while the
warrior in purple stands firm and executes a sweeping counter-attack with a
flaming effect. They clash in the center of the frame, creating a shockwave of
energy and sparks. The camera rapidly orbits around them as the fiery battle
intensifies, concluding with the red-clad swordsman landing gracefully while the
purple warrior points his weapon forward, both facing the camera.
Environment: Dramatic background with embers floating in the air, high contrast
lighting.
</code></pre>
<h2>4. 坑与解法（现象/根因/解法）</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>现象</th>
<th>根因</th>
<th>解法</th>
</tr>
</thead>
<tbody>
<tr>
<td>参考图完全不生效，出片成纯文生视频，无任何报错</td>
<td>AUTOGROW 输入按 dict 消费（源码 <code>(ref_images or {}).values()</code>），/prompt 传 list 被静默丢弃</td>
<td>扁平点号键：<code>"ref_images.ref_image_0": ["&lt;LoadImage节点id&gt;", 0]</code>（已用官方 execution 测试用例验证该格式）</td>
</tr>
<tr>
<td>289 帧解码阶段 <code>hipErrorIllegalAddress</code>，随后队列全部任务连环瞬崩</td>
<td>ROCm 分配失败直接毒化进程级 GPU 上下文；ComfyUI 模型跨任务常驻缓存，上任务残留 + 下任务换装叠加爆卡</td>
<td>重启 ComfyUI 清上下文；长任务前先重启清缓存；队列串行 ≠ 显存安全</td>
</tr>
<tr>
<td>朋友侧工作流在 GPU 上 spawn 独立 llama-server（17.5G），与 H3 模型叠加必炸</td>
<td>提示词节点的 llama-server 绕过 ComfyUI 显存管理器</td>
<td>双卡实例隔离：专用卡不对外，对外卡独立 ComfyUI 进程</td>
</tr>
<tr>
<td>穿透访问启动页卡 3 分钟+</td>
<td>object_info 未压缩 2.7MB 过跨洋隧道（13KB/s）</td>
<td>nginx 反代 gzip：2.7MB→324KB，206s→9.6s；静态资产加 30d immutable 缓存</td>
</tr>
</tbody>
</table>
<h2>5. 实测数据</h2>
<p dir="auto"><strong>单卡正式出片</strong>（1344×768 / 294 帧 / 20 步 / 无 LoRA / seed 12345 / 6 参考图 match）：</p>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>版本</th>
<th>journal Total Time</th>
<th>规格</th>
</tr>
</thead>
<tbody>
<tr>
<td>3D CGI</td>
<td>4650.3 s（77.5 min）</td>
<td>1344×768 / 12.25s / 294 帧 / 24fps</td>
</tr>
<tr>
<td>2D 手绘</td>
<td>4590.3 s（76.5 min）</td>
<td>同上</td>
</tr>
</tbody>
</table>
<p dir="auto"><strong>同规格提速对照</strong>（1344×768 / 294 帧）：</p>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>配置</th>
<th>耗时</th>
<th>画质</th>
</tr>
</thead>
<tbody>
<tr>
<td>ref2v_turbo_4step LoRA（4 步）</td>
<td>870 s（14.5 min）</td>
<td>细节软，角色可辨</td>
</tr>
<tr>
<td>无 LoRA 20 步（参赛档）</td>
<td>4590~4650 s</td>
<td>全质量</td>
</tr>
</tbody>
</table>
<p dir="auto"><strong>注意力机制横评</strong>（864×480 / 124 帧 / 4 步 / 同 seed / 单卡独占）：</p>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>模式</th>
<th>总耗时</th>
</tr>
</thead>
<tbody>
<tr>
<td>PyTorch SDPA（默认）</td>
<td>195.2 s</td>
</tr>
<tr>
<td>comfy_kitchen int8 attention（--use-ck-attention）</td>
<td>195.1 s</td>
</tr>
<tr>
<td>SageAttention 3.3.12</td>
<td>210.2 s</td>
</tr>
</tbody>
</table>
<p dir="auto">三种模式逐帧 PSNR 11~13 dB（轨迹分叉，4 步蒸馏模型对数值扰动敏感），画面均无伪影。<strong>结论：量化内核路线下 attention 非瓶颈，默认 SDPA 即可，Sage 负优化。</strong></p>
<h2>6. 证据清单</h2>
<ul>
<li>完整日志：ComfyUI journal（got prompt → Prompt executed 全程），随时可贴</li>
<li>工作流：API JSON（§3 结构 + §4 扁平键格式）可直接复现</li>
<li>视频：
<ul>
<li><strong>3D CGI 版</strong>：<a href="https://youtu.be/7dSj4cW1PaE" rel="nofollow ugc">https://youtu.be/7dSj4cW1PaE</a></li>
<li><strong>2D 手绘版</strong>：<a href="https://youtu.be/cLiOQ2skN5Q" rel="nofollow ugc">https://youtu.be/cLiOQ2skN5Q</a></li>
</ul>
</li>
</ul>
<p dir="auto">硬件：X99-TF / E5-2696 v3 / 128G ECC / 2× R9700 32G / ROCm 7.2.4。</p>
<p dir="auto">有同卡（gfx1201）跑 H3 的欢迎交流，重点碰两个问题：AUTOGROW 的 API 格式、294 帧解码的显存尖峰。</p>
]]></description><link>https://lcz.me/topic/3139</link><generator>RSS for Node</generator><lastBuildDate>Thu, 19 Nov 2026 00:16:54 GMT</lastBuildDate><atom:link href="https://lcz.me/topic/3139.rss" rel="self" type="application/rss+xml"/><pubDate>Tue, 17 Nov 2026 02:15:49 GMT</pubDate><ttl>60</ttl><item><title><![CDATA[Reply to MiniMax H3 视频大赛参赛帖：双 R9700 本地管线 1344×768/12s 全质量出片 on Tue, 17 Nov 2026 07:00:36 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/bunsei" aria-label="Profile: Bunsei">@<bdi>Bunsei</bdi></a> 已经测完单卡，速度提升一倍，<img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/1f44d.png?v=3c59dda5525" class="not-responsive emoji emoji-android emoji--+1" style="height:1.4em;width:auto;vertical-align:middle" title=":+1:" alt="👍" /></p>
<p dir="auto">单发测试帖子了：<a href="https://lcz.me/topic/3144">https://lcz.me/topic/3144</a></p>
]]></description><link>https://lcz.me/post/34083</link><guid isPermaLink="true">https://lcz.me/post/34083</guid><dc:creator><![CDATA[风满楼]]></dc:creator><pubDate>Tue, 17 Nov 2026 07:00:36 GMT</pubDate></item><item><title><![CDATA[Reply to MiniMax H3 视频大赛参赛帖：双 R9700 本地管线 1344×768/12s 全质量出片 on Tue, 17 Nov 2026 06:11:36 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/aaronku" aria-label="Profile: aaronku">@<bdi>aaronku</bdi></a> 站里大佬很多，我也刚配起来跑，还在学习阶段。R9700 双卡可以单卡一张卡跑 LLM，一张卡跑 comfyui，也可以双卡 TP 跑大模型和 comfyui。</p>
<p dir="auto">AI MAX395 应该也一样的道理，不过 395 应该没有告诉互联的办法吧，我猜测，单台跑 LLM，生图生视频应该都还是跑的动的。手头没有机器只能瞎猜。</p>
<p dir="auto">自己玩不转就让 agent 帮你，workbuddy 可以搞，hermes，claude code， codex 等等任何一个都可以拿来帮你驱动你的机器。</p>
<p dir="auto">互相学习<img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/1f44d.png?v=3c59dda5525" class="not-responsive emoji emoji-android emoji--+1" style="height:1.4em;width:auto;vertical-align:middle" title=":+1:" alt="👍" /></p>
]]></description><link>https://lcz.me/post/34071</link><guid isPermaLink="true">https://lcz.me/post/34071</guid><dc:creator><![CDATA[风满楼]]></dc:creator><pubDate>Tue, 17 Nov 2026 06:11:36 GMT</pubDate></item><item><title><![CDATA[Reply to MiniMax H3 视频大赛参赛帖：双 R9700 本地管线 1344×768/12s 全质量出片 on Tue, 17 Nov 2026 04:09:35 GMT]]></title><description><![CDATA[<p dir="auto">太糊了，通过初审，20积分比赛结束后统一结算，不加优秀。</p>
]]></description><link>https://lcz.me/post/34046</link><guid isPermaLink="true">https://lcz.me/post/34046</guid><dc:creator><![CDATA[terry]]></dc:creator><pubDate>Tue, 17 Nov 2026 04:09:35 GMT</pubDate></item><item><title><![CDATA[Reply to MiniMax H3 视频大赛参赛帖：双 R9700 本地管线 1344×768/12s 全质量出片 on Tue, 17 Nov 2026 03:11:35 GMT]]></title><description><![CDATA[<p dir="auto">站长让我来跟你学习 学习，我有两张 R9700 跟两台AI MAX395  不知道能做什么~</p>
]]></description><link>https://lcz.me/post/34028</link><guid isPermaLink="true">https://lcz.me/post/34028</guid><dc:creator><![CDATA[aaronku]]></dc:creator><pubDate>Tue, 17 Nov 2026 03:11:35 GMT</pubDate></item><item><title><![CDATA[Reply to MiniMax H3 视频大赛参赛帖：双 R9700 本地管线 1344×768/12s 全质量出片 on Tue, 17 Nov 2026 02:39:09 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/%E9%A3%8E%E6%BB%A1%E6%A5%BC" aria-label="Profile: 风满楼">@<bdi>风满楼</bdi></a> 安装和配置稍微有点麻烦，不过可以丢给AI Agent 来让他帮你配。</p>
]]></description><link>https://lcz.me/post/34012</link><guid isPermaLink="true">https://lcz.me/post/34012</guid><dc:creator><![CDATA[Bunsei]]></dc:creator><pubDate>Tue, 17 Nov 2026 02:39:09 GMT</pubDate></item><item><title><![CDATA[Reply to MiniMax H3 视频大赛参赛帖：双 R9700 本地管线 1344×768/12s 全质量出片 on Tue, 17 Nov 2026 02:30:30 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/bunsei" aria-label="Profile: Bunsei">@<bdi>Bunsei</bdi></a> 好，一会试试</p>
]]></description><link>https://lcz.me/post/34010</link><guid isPermaLink="true">https://lcz.me/post/34010</guid><dc:creator><![CDATA[风满楼]]></dc:creator><pubDate>Tue, 17 Nov 2026 02:30:30 GMT</pubDate></item><item><title><![CDATA[Reply to MiniMax H3 视频大赛参赛帖：双 R9700 本地管线 1344×768/12s 全质量出片 on Tue, 17 Nov 2026 02:23:36 GMT]]></title><description><![CDATA[<p dir="auto">可以试试这套节点不？想知道在咱们的这套配置上速度大概能提高多少。 <a href="https://codeberg.org/Bunsei/ComfyUI-H3-RDNA4/releases/tag/v0.2.0" rel="nofollow ugc">Codeberg</a></p>
]]></description><link>https://lcz.me/post/34008</link><guid isPermaLink="true">https://lcz.me/post/34008</guid><dc:creator><![CDATA[Bunsei]]></dc:creator><pubDate>Tue, 17 Nov 2026 02:23:36 GMT</pubDate></item></channel></rss>