<?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] Mac Studio M5 Max 40 GPU 64 GB 统一内存｜MiniMax H3 3D / 2D ｜Draw Things 单卡一次跑完 294 帧]]></title><description><![CDATA[<h2>前言</h2>
<p dir="auto">重在参与。用的Drawthings, 没有调教，没有任何优化。</p>
<p dir="auto">这次参赛没有分段、没有 LoRA、没有 turbo 加速，也没有 ffmpeg 拼接。两条 12 秒成片都是同一台 Mac Studio 上、同一个 Draw Things 项目里，一次生成 294 帧（1344×768）出来的。下面所有参数、耗时都来自 Draw Things 项目数据库里的计时 log（每条生成记录的 <code>duration</code> / <code>timings</code>），不是估算。</p>
<ul>
<li>3D CGI 成片：<a href="https://youtu.be/7X3QUBg-3LQ" rel="nofollow ugc">https://youtu.be/7X3QUBg-3LQ</a></li>
<li>传统 2D 动画成片：<a href="https://youtu.be/noVovfnEz7s" rel="nofollow ugc">https://youtu.be/noVovfnEz7s</a></li>
</ul>
<hr />
<h2>硬件 / 软件环境</h2>
<ul>
<li>机器：Mac Studio</li>
<li>芯片：Apple M5 Max，18 核（6 超级核 + 12 性能核）</li>
<li>内存：64 GB 统一内存</li>
<li>OS：macOS 27.0.1</li>
<li>推理栈：Draw Things 26.1020.0（CoreML / Metal，纯苹果栈，没有 ComfyUI、没有 ROCm、没有 CUDA）</li>
<li>输出格式：Draw Things 原生视频 + 同步音频 → ProRes 4444（1344×768, 24 fps）+ PCM 32 kHz 立体声</li>
</ul>
<p dir="auto">M 卡跑 H3 的难点不是"跑不通"，而是"跑得慢"：294 帧整段生成，采样阶段占了 98% 以上的时间，单步 4 分钟上下。能一次跑完 294 帧靠的是下面几个设定。</p>
<h2>模型档案</h2>
<ul>
<li><code>minimax_h3_ref2va_i8x.ckpt</code> — 34.4 GB（Ref2VA 基座，i8x 量化）</li>
<li><code>qwen_3_vl_32b_50_i8x.ckpt</code> — 26.5 GB（Qwen3-VL-32B 文字编码器，i8x）</li>
<li><code>minimax_h3_vae_f16.ckpt</code> — 5.51 GB（视频 / 音频 VAE，FP16）</li>
<li>合计约 66.4 GB 权重，全部在 64 GB 统一内存里跑</li>
<li>LoRA：无（<code>loras: []</code>）｜upscaler：无｜face restoration：无｜hiresFix：关闭</li>
<li>参考图走 canvas / moodboard 喂入（比赛主帖指定的那组剑心、志志雄参考图）</li>
</ul>
<h2>关键参数（两条片子除 Style 与 seed 之外完全一致）</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>项</th>
<th>3D 版</th>
<th>2D 版</th>
</tr>
</thead>
<tbody>
<tr>
<td>模型</td>
<td><code>minimax_h3_ref2va_i8x.ckpt</code></td>
<td>同</td>
</tr>
<tr>
<td>分辨率</td>
<td>1344 × 768</td>
<td>1344 × 768</td>
</tr>
<tr>
<td>帧数</td>
<td>294</td>
<td>294</td>
</tr>
<tr>
<td>步数</td>
<td>50</td>
<td>50</td>
</tr>
<tr>
<td>采样器</td>
<td>DDIM Trailing（<code>sampler: 16</code>）</td>
<td>同</td>
</tr>
<tr>
<td>Shift / Audio Shift</td>
<td>12 / 3</td>
<td>12 / 3</td>
</tr>
<tr>
<td>Text guidance</td>
<td>1</td>
<td>1</td>
</tr>
<tr>
<td>Seed</td>
<td>2254839087</td>
<td>1694873285</td>
</tr>
<tr>
<td>Seed mode</td>
<td>ScaleAlike（<code>seedMode: 2</code>）</td>
<td>同</td>
</tr>
<tr>
<td>分块解码</td>
<td>开启，tile 256×256，overlap 64</td>
<td>同</td>
</tr>
<tr>
<td>分块扩散</td>
<td>关闭</td>
<td>关闭</td>
</tr>
<tr>
<td>teaCache</td>
<td>关闭</td>
<td>关闭</td>
</tr>
<tr>
<td>maskBlur / maskBlurOutset</td>
<td>1.5 / 0</td>
<td>1.5 / 0</td>
</tr>
<tr>
<td>sharpness</td>
<td>0</td>
<td>0</td>
</tr>
<tr>
<td>batchCount</td>
<td>1</td>
<td>1</td>
</tr>
<tr>
<td>strength</td>
<td>1</td>
<td>1</td>
</tr>
</tbody>
</table>
<p dir="auto">这些数值就是官方 Ref2VA 的推荐起点（768 短边、50 步 DDIM Trailing、Text Guidance 1、Shift 12、Audio Shift 3），没有改动。</p>
<h2>为什么 64 GB 能一次跑完 294 帧</h2>
<ol>
<li><strong>i8x 量化基座 + i8x 文字编码器</strong>：34.4 GB + 26.5 GB，这是 Draw Things 官方给 H3 的量化版本，也是 64 GB 机器唯一可行的组合。</li>
<li><strong><code>tiledDecoding: true</code>（tile 256×256 / overlap 64）</strong>：VAE 解码按块进行，294 帧不是一次性解码，解码阶段只用了 122.6 s / 117.0 s。</li>
<li><strong><code>tiledDiffusion: false</code></strong>：扩散本身整块跑 —— 说明 294 帧的 latent 在 64 GB 里放得下，不需要分段。</li>
<li><strong>没有 turbo LoRA</strong>：官方 Ref2VA 基座没有 4-step turbo 分支，所以只能 50 步硬跑，这是耗时的根源。</li>
</ol>
<h2>实测耗时（取自项目 sqlite 里的 <code>duration</code> / <code>timings</code>）</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>版本</th>
<th>总耗时</th>
<th>text_encoded</th>
<th>controls_generated</th>
<th>sampling</th>
<th>image_decoded</th>
</tr>
</thead>
<tbody>
<tr>
<td>3D</td>
<td>12,542.3 s（3 h 29 m 02 s）</td>
<td>43.8 s</td>
<td>22.4 s</td>
<td>12,353.2 s（98.5%）</td>
<td>122.6 s</td>
</tr>
<tr>
<td>2D</td>
<td>13,619.0 s（3 h 46 m 59 s）</td>
<td>65.8 s</td>
<td>5.9 s</td>
<td>13,430.1 s（98.6%）</td>
<td>117.0 s</td>
</tr>
</tbody>
</table>
<p dir="auto">采样阶段逐步耗时：</p>
<ul>
<li>3D：首条 log 22.6 s（首块预热），其余 50 步平均 <strong>246.6 s/步</strong>，区间 238.3 – 254.4 s</li>
<li>2D：首条 log 14.7 s，其余 50 步平均 <strong>268.3 s/步</strong>，区间 259.9 – 279.4 s</li>
</ul>
<p dir="auto">2D 版比 3D 版慢 1,076.7 s（约 18 分钟），每步慢约 22 s。</p>
<h2>成片校验</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>项</th>
<th>3D</th>
<th>2D</th>
</tr>
</thead>
<tbody>
<tr>
<td>分辨率</td>
<td>1344 × 768</td>
<td>1344 × 768</td>
</tr>
<tr>
<td>帧率 / 帧数</td>
<td>24 fps / 294 帧</td>
<td>24 fps / 294 帧</td>
</tr>
<tr>
<td>时长</td>
<td>12.25 s</td>
<td>12.25 s</td>
</tr>
<tr>
<td>视频编码</td>
<td>ProRes 4444（ap4h）</td>
<td>ProRes 4444（ap4h）</td>
</tr>
<tr>
<td>音频</td>
<td>PCM 32 kHz 立体声，非静音</td>
<td>PCM 32 kHz 立体声，非静音</td>
</tr>
<tr>
<td>文件大小</td>
<td>326 MB</td>
<td>305 MB</td>
</tr>
</tbody>
</table>
<p dir="auto">直接在 1344×768 采样，无缩放、无补帧、无裁切。</p>
<h2>Prompt（统一英文，只切 Style）</h2>
<p dir="auto">3D 版：</p>
<pre><code>Style: 3D CGI animation  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>
<p dir="auto">2D 版：开头换成 <code>2D traditional anime</code>，其余一字不差。</p>
<h2>画面对照</h2>
<p dir="auto">两版都完成了 prompt 要求的：红衣剑客突进、火焰横扫反击、中央交锋、镜头斜向/低角度环绕、结尾双方持刃相对。</p>
<ul>
<li><strong>3D 版</strong>：材质光滑、光影立体，人物接近手办/三维渲染质感；火星与火焰柱构成包围式构图，轮廓光明显。</li>
<li><strong>2D 版</strong>：线条清晰、色块分明、阴影戏剧化，典型赛璐璐手绘动画调性；剑刃交错与橙黄光带形成对角线构图，红发紫眼剑客与缠脸紫袍对手的对决更"动画化"。</li>
</ul>
<h2>复现说明</h2>
<p dir="auto">附件 <code>3D.json</code> / <code>2D.json</code> 就是这两条成片对应的 Draw Things 生成参数导出，逐字段可核对。同环境（M5 Max 64 GB + Draw Things 26.1020.0 + 上述三个权重）用同一 prompt、同一 seed 重跑即可复现：3D seed 2254839087，2D seed 1694873285。</p>
<p dir="auto">需要留意的两点：</p>
<ol>
<li>64 GB 是这条链路的下限体验，294 帧整段扩散靠 i8x 量化才放得下；换 i6x 或加 hiresFix 会撑爆。</li>
<li>没有 4-step turbo LoRA 可用（Ref2VA 基座不支持），所以 50 步是硬成本，一条 12 秒成片约 3.5 小时。</li>
</ol>
<p dir="auto"><a href="https://upload.lcz.me/uploads/22a08a5e-f4e5-48d7-bffe-2ed2e017f30c.json" rel="nofollow ugc">2D.json</a> <a href="https://upload.lcz.me/uploads/b49d560e-f6a8-4b41-8e89-ea55aa174f3f.json" rel="nofollow ugc">3D.json</a></p>
]]></description><link>https://lcz.me/topic/3160</link><generator>RSS for Node</generator><lastBuildDate>Thu, 19 Nov 2026 00:16:38 GMT</lastBuildDate><atom:link href="https://lcz.me/topic/3160.rss" rel="self" type="application/rss+xml"/><pubDate>Tue, 17 Nov 2026 20:33:33 GMT</pubDate><ttl>60</ttl><item><title><![CDATA[Reply to [H3剑心比赛] Mac Studio M5 Max 40 GPU 64 GB 统一内存｜MiniMax H3 3D / 2D ｜Draw Things 单卡一次跑完 294 帧 on Wed, 18 Nov 2026 00:41:46 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> 确实拉，后面剑心的剑还穿模了 <img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/1f630.png?v=3c59dda5525" class="not-responsive emoji emoji-android emoji--cold_sweat" style="height:1.4em;width:auto;vertical-align:middle" title=":cold_sweat:" alt="😰" /></p>
]]></description><link>https://lcz.me/post/34237</link><guid isPermaLink="true">https://lcz.me/post/34237</guid><dc:creator><![CDATA[fsclsnow]]></dc:creator><pubDate>Wed, 18 Nov 2026 00:41:46 GMT</pubDate></item><item><title><![CDATA[Reply to [H3剑心比赛] Mac Studio M5 Max 40 GPU 64 GB 统一内存｜MiniMax H3 3D / 2D ｜Draw Things 单卡一次跑完 294 帧 on Tue, 17 Nov 2026 20:42:58 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 e3160-412-1</p>
</blockquote>
]]></description><link>https://lcz.me/post/34213</link><guid isPermaLink="true">https://lcz.me/post/34213</guid><dc:creator><![CDATA[Xiaote]]></dc:creator><pubDate>Tue, 17 Nov 2026 20:42:58 GMT</pubDate></item><item><title><![CDATA[Reply to [H3剑心比赛] Mac Studio M5 Max 40 GPU 64 GB 统一内存｜MiniMax H3 3D / 2D ｜Draw Things 单卡一次跑完 294 帧 on Tue, 17 Nov 2026 20:42:54 GMT]]></title><description><![CDATA[<p dir="auto">2D版本运镜太拉了，整体过审，20积分比赛结束后统一结算。数据很值钱，缺这个机型，破格打个优秀。</p>
]]></description><link>https://lcz.me/post/34212</link><guid isPermaLink="true">https://lcz.me/post/34212</guid><dc:creator><![CDATA[terry]]></dc:creator><pubDate>Tue, 17 Nov 2026 20:42:54 GMT</pubDate></item></channel></rss>