<?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】M5 Ultra 256G H3.C引擎跑MiniMax H3模型生成2D/3D 12秒视频]]></title><description><![CDATA[<h1>Mac Studio 上用 h3.c 原生引擎生成两条 12 秒 参赛视频，Apple M5 Ultra</h1>
<blockquote>
<p dir="auto">MiniMax H3 视频模型，294 帧 / 1344×768 / 带音效，单卷 37 分钟出片。<br />
两个参赛作品，同一个引擎、同一套参数、提示词逐字节未改。</p>
</blockquote>
<hr />
<h2>🧭 这是什么</h2>
<p dir="auto">MiniMax H3 的官方权重是 BF16 的，社区跑法大多走 Python 移植或量化转换。这篇用的是 <strong>h3.c</strong>——antirez（Redis 作者）写的原生 C/Metal 引擎，没有 Python 运行时，权重 mmap 直接进 Metal buffer。</p>
<p dir="auto">两条视频分别是：</p>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th></th>
<th>作品一</th>
<th>作品二</th>
</tr>
</thead>
<tbody>
<tr>
<td>风格</td>
<td>2D 手绘动画对战</td>
<td>3D CGI 对战</td>
</tr>
<tr>
<td>条件</td>
<td>双参考图（Ref2VA）</td>
<td>双参考图（Ref2VA）</td>
</tr>
<tr>
<td>成片</td>
<td>294 帧 / 12.25 s / 37.6 MB</td>
<td>294 帧 / 12.25 s / 36.4 MB</td>
</tr>
</tbody>
</table>
<hr />
<h2><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/1f5a5.png?v=3c59dda5525" class="not-responsive emoji emoji-android emoji--desktop_computer" style="height:1.4em;width:auto;vertical-align:middle" title="🖥" alt="🖥" /> 测试环境</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>项</th>
<th>值</th>
</tr>
</thead>
<tbody>
<tr>
<td>机型</td>
<td>Mac Studio</td>
</tr>
<tr>
<td>芯片</td>
<td>Apple M5 Ultra，36 核（12 Super + 24 Performance）</td>
</tr>
<tr>
<td>统一内存</td>
<td>256 GB</td>
</tr>
<tr>
<td>GPU</td>
<td>Metal 4 TensorOps，单 buffer 上限 180 GiB</td>
</tr>
<tr>
<td>推理引擎</td>
<td>h3.c（h3-metal 0.1.0-dev，commit <code>8974cc0</code>，编码器补丁见下文）</td>
</tr>
<tr>
<td>权重</td>
<td>官方 <code>MiniMaxAI/MiniMax-H3</code> BF16 快照（DiT 61.73 GiB + Qwen3-VL 编码器 62.13 GiB + video VAE 9.70 GiB + audio VAE 0.56 GiB），加载时 int8 量化</td>
</tr>
</tbody>
</table>
<hr />
<h2><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/1f6e0.png?v=3c59dda5525" class="not-responsive emoji emoji-android emoji--hammer_and_wrench" style="height:1.4em;width:auto;vertical-align:middle" title="🛠" alt="🛠" /> 抄作业步骤</h2>
<p dir="auto">两条命令，逐字可粘贴（<code>-d</code> 指权重目录，<code>-p</code> 换成你自己的提示词文件）：</p>
<p dir="auto"><strong>作品一（2D，双参考图）</strong></p>
<pre><code class="language-bash">./h3 --profile -d ~/h3/MiniMax-H3-R2 \
  -p "$(cat prompt_2d.txt)" \
  --seconds 12 --width 1344 --height 768 \
  --steps 20 --reuse 2 --seed 42 \
  --ref-image ~/refs/560.jpeg \
  --ref-image ~/refs/紫袍v2.jpeg \
  -o h3c_contest_2d_r2_1344x768_12s_crf10.mp4
</code></pre>
<p dir="auto"><strong>作品二（3D，双参考图）</strong></p>
<pre><code class="language-bash">./h3 --profile -d ~/h3/MiniMax-H3-R2 \
  -p "$(cat prompt_3d.txt)" \
  --seconds 12 --width 1344 --height 768 \
  --steps 20 --reuse 2 --seed 42 \
  --ref-image ~/refs/红袍.jpeg \
  --ref-image ~/refs/紫袍.jpeg \
  -o h3c_contest_r2_1344x768_12s_crf10.mp4
</code></pre>
<p dir="auto">参考图与提示词的哈希（证明发布件与跑片件逐字节一致）：</p>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>资产</th>
<th>尺寸</th>
<th>sha256</th>
</tr>
</thead>
<tbody>
<tr>
<td>560.jpeg（作品一参考 1，剑心）</td>
<td>1378×2821</td>
<td><code>1af483d0…27b9036f</code></td>
</tr>
<tr>
<td>紫袍v2.jpeg（作品一参考 2，志志雄）</td>
<td>564×562</td>
<td><code>30137b10…5eab4b4f</code></td>
</tr>
<tr>
<td>红袍.jpeg（作品二参考 1）</td>
<td>976×1456</td>
<td><code>15259ece…653e67d6</code></td>
</tr>
<tr>
<td>紫袍.jpeg（作品二参考 2）</td>
<td>550×800</td>
<td><code>0c9baafd…749bcb17</code></td>
</tr>
<tr>
<td>prompt_2d.txt</td>
<td>—</td>
<td><code>eef33f0a…9ec1c737</code></td>
</tr>
<tr>
<td>prompt_3d.txt</td>
<td>—</td>
<td><code>2331bac1…db270bb</code></td>
</tr>
</tbody>
</table>
<p dir="auto">时长口径：H3 帧数必须落在 24 fps 的 <code>5+17n</code> 网格上，12 秒（288 帧）向上吸附到 294 帧 = 12.25 秒。ffprobe 实测两个文件均为 <code>duration=12.250000</code>、<code>nb_frames=294</code>。</p>
<hr />
<h2>🧬 JSON 工作流</h2>
<p dir="auto">作品一（Ref2VA 双参考图，2D 动漫）：</p>
<pre><code class="language-json">{
  "engine": "h3.c native C/Metal, commit 8974cc0, encoder patch CRF10, binary sha256 cfc3e3052d7a90fe60fdf5446498e5f533b2b6f76eb428a81b1c19f5b77b8d40",
  "model": "MiniMaxAI/MiniMax-H3 BF16 snapshot, Ref2VA partition, int8 quantized at load",
  "command": ["./h3","--profile","-d","~/h3/MiniMax-H3-R2",
    "-p","&lt;PROMPT_TXT verbatim, sha256 eef33f0a3c7a849bdde0a23a08d5f20e98a02532b8c9b43466d4dfae9ec1c737&gt;",
    "--seconds","12","--width","1344","--height","768",
    "--steps","20","--reuse","2","--seed","42",
    "--ref-image","~/refs/560.jpeg (1378x2821, sha256 1af483d0813e92b87108083d7e68f46afdbcb58ea1b4ac59f4fcdd2327b9036f)",
    "--ref-image","~/refs/紫袍v2.jpeg (564x562, sha256 30137b10d9530c4435994791eda21fa4b441143f25f4526a3e03ef9a5eab4b4f)",
    "-o","h3c_contest_2d_r2_1344x768_12s_crf10.mp4"],
  "canvas": "1344x768",
  "frames": "294 (12.25s @24fps; H3 grid 5+17n, 12s rounds up)",
  "steps": 20, "reuse": 2, "layers": 50, "seed": 42,
  "conditioning": "Ref2VA ordered references: Picture 1 = 560 (Jianxin), Picture 2 = 紫袍v2 (Zhizhixiong)",
  "audio": "model-generated, 32kHz stereo AAC",
  "output": "h3c_contest_2d_r2_1344x768_12s_crf10.mp4, 37,622,992 B, sha256 770eaa91f59aaa51a1815697b750825cac1e1a8beb5ccd01d4d4f7617003b439"
}
</code></pre>
<p dir="auto">作品二（Ref2VA 双参考图，3D CGI）：</p>
<pre><code class="language-json">{
  "engine": "h3.c native C/Metal, commit 8974cc0, encoder patch CRF10, binary sha256 cfc3e3052d7a90fe60fdf5446498e5f533b2b6f76eb428a81b1c19f5b77b8d40",
  "model": "MiniMaxAI/MiniMax-H3 BF16 snapshot, Ref2VA partition, int8 quantized at load",
  "command": ["./h3","--profile","-d","~/h3/MiniMax-H3-R2",
    "-p","&lt;PROMPT_TXT verbatim, sha256 2331bac16efd4a8183a9499ef01eaff9b095eef0eaa4a34f1f9ccfe55db270bb&gt;",
    "--seconds","12","--width","1344","--height","768",
    "--steps","20","--reuse","2","--seed","42",
    "--ref-image","~/refs/红袍.jpeg (976x1456, sha256 15259ece82370268a1002cab252b0e846cc7871352ed83d930ecb464653e67d6)",
    "--ref-image","~/refs/紫袍.jpeg (550x800, sha256 0c9baafdd016ee8952569783947fe430ac872398f9ed6cddfc093de3749bcb17)",
    "-o","h3c_contest_r2_1344x768_12s_crf10.mp4"],
  "canvas": "1344x768",
  "frames": "294 (12.25s @24fps)",
  "steps": 20, "reuse": 2, "layers": 50, "seed": 42,
  "conditioning": "Ref2VA ordered references: Picture 1 = 红袍 (Jianxin), Picture 2 = 紫袍 (Zhizhixiong)",
  "audio": "model-generated, 32kHz stereo AAC",
  "output": "h3c_contest_r2_1344x768_12s_crf10.mp4, 36,393,959 B, sha256 6721794f2f3094ae6129733bef396bef270fd16b45ea65ecbb661a996dbba84c"
}
</code></pre>
<hr />
<h2><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/23f1.png?v=3c59dda5525" class="not-responsive emoji emoji-android emoji--stopwatch" style="height:1.4em;width:auto;vertical-align:middle" title="⏱" alt="⏱" /> 实测成绩</h2>
<p dir="auto">单卷独跑基线（<code>--profile</code> 分段计时，文件系统时间戳核验）：</p>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>阶段</th>
<th>作品一</th>
<th>作品二</th>
</tr>
</thead>
<tbody>
<tr>
<td>条件编码</td>
<td>≈24 s</td>
<td>≈19 s</td>
</tr>
<tr>
<td>DiT 加载（mmap）</td>
<td>24.2 s</td>
<td>18.5 s</td>
</tr>
<tr>
<td><strong>DiT 去噪（20 步，11 次全量评估）</strong></td>
<td><strong>1989.8 s</strong></td>
<td><strong>1999.3 s</strong></td>
</tr>
<tr>
<td>audio VAE 解码</td>
<td>6.6 s</td>
<td>6.6 s</td>
</tr>
<tr>
<td>video VAE 解码</td>
<td>199.3 s</td>
<td>183.8 s</td>
</tr>
<tr>
<td><strong>端到端</strong></td>
<td><strong>37 分 29 秒</strong></td>
<td><strong>37 分 17 秒</strong></td>
</tr>
</tbody>
</table>
<p dir="auto">生效的加速手段（全部可在日志对账）：</p>
<ul>
<li><strong>int8 MLP / QKV / 注意力输出投影</strong>（M5 默认路径）：量化完成后释放 BF16 权重，峰值张量存储 32.2 GiB（BF16 路径 36.4 GiB）</li>
<li><strong>核融合</strong>：注意力残差门控 + MLP AdaLN、跨块 AdaLN 携带、SwiGLU 单图执行</li>
<li><strong><code>--reuse 2</code> 速度复用日程</strong>：20 步只算 11 次完整去噪（日志 <code>selected GPU reuse schedule has 11 evaluations</code>）</li>
<li><strong>零拷贝权重映射</strong>：safetensors 直接 mmap，61.7 GiB DiT 加载仅 18.5–32.2 s</li>
<li><strong>编码器补丁</strong>：x264 出厂 CRF 18 → CRF 10（一行改动重编译），火花/火焰纹理保留完整</li>
</ul>
<p dir="auto">未启用（如实登记）：ANE 卸载（本机实测每步 -30%）、ssd-streaming、token-reduction、层削减。</p>
<hr />
<h2><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/1f4cb.png?v=3c59dda5525" class="not-responsive emoji emoji-android emoji--clipboard" style="height:1.4em;width:auto;vertical-align:middle" title="📋" alt="📋" /> 已知瑕疵：16px 网格纹理</h2>
<p dir="auto">放大看成片，背景大面积区域有一层 16 像素周期的网格纹理。这不是我这条链路的个例：</p>
<ul>
<li>ComfyUI 社区 issue <a href="https://github.com/Comfy-Org/ComfyUI/issues/15416" rel="nofollow ugc">#15416</a> 对 H3 视频 VAE 做过量化测量：整图解码误差 93.7 vs 分块 31.4，症状描述一致，且与量化位宽无关</li>
<li>我把能试的都试了一遍：int8 / BF16 全量、VAE 分块参数、去块滤镜、30 步 + 每步全算的参考质量档——网格幅度有轻有重，<strong>没有一个配置能消掉它</strong></li>
<li>结论：这是模型本身的输出纹理，量化与后处理只能改幅度。仙宫的生成视频（更早的小分辨率试跑）不明显，是因为 864×480 下每个 16px 格子占画面比例小，不是配置差异</li>
</ul>
<p dir="auto">这个分辨率下生成视频网格是可见的，接受它或者降低输出分辨率。</p>
<hr />
<h2><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/270d.png?v=3c59dda5525" class="not-responsive emoji emoji-android emoji--writing_hand" style="height:1.4em;width:auto;vertical-align:middle" title="✍" alt="✍" />️ 提示词</h2>
<p dir="auto">两条提示词原文投喂、零改动，哈希见上表。</p>
<p dir="auto"><strong>作品一</strong>（<code>eef33f0a…9ec1c737</code>）：</p>
<blockquote>
<p dir="auto">2D traditional anime) Masterpiece, high quality, cinematic lighting.<br />
Subject: A fierce battle sequence featuring two characters based on the reference images—a red-clad swordsman (Jianxin) and a purple-robed warrior (Zhizhixiong).<br />
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.<br />
Environment: Dramatic background with embers floating in the air, high contrast lighting.</p>
</blockquote>
<p dir="auto"><strong>作品二</strong>（<code>2331bac1…db270bb</code>）：</p>
<blockquote>
<p dir="auto">Style: 3D CGI animation Masterpiece, high quality, cinematic lighting.<br />
Subject: A fierce battle sequence featuring two characters based on the reference images—a red-clad swordsman (Jianxin) and a purple-robed warrior (Zhizhixiong).<br />
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.<br />
Environment: Dramatic background with embers floating in the air, high contrast lighting.</p>
</blockquote>
<hr />
<h2><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/1f3ac.png?v=3c59dda5525" class="not-responsive emoji emoji-android emoji--clapper" style="height:1.4em;width:auto;vertical-align:middle" title="🎬" alt="🎬" /> 参赛作品</h2>
<p dir="auto"><strong>生成视频一：2D 动漫双人对战</strong></p>
<p dir="auto"><a href="https://youtu.be/jg7mdJC0ha8" rel="nofollow ugc"><i class="fa fa-youtube" aria-hidden="true"></i> Youtube Video</a></p><div class="js-lazyYT lazyYT-container" data-youtube-id="jg7mdJC0ha8" data-width="640" data-height="360" data-parameters style="width:640px;padding-bottom:360px">
 <div class="ytp-thumbnail lazyYT-image-loaded" style="background-image:url(&quot;https://i.ytimg.com/vi/jg7mdJC0ha8/hqdefault.jpg&quot;)">
  <button class="ytp-large-play-button ytp-button" tabindex="23" aria-live="assertive" style="transform:scale(0.85)" onclick="$(this).lazyYT(this);return false;">
   <svg height="100%" version="1.1" viewbox="0 0 68 48" width="100%">
    <path class="ytp-large-play-button-bg" d="m .66,37.62 c 0,0 .66,4.70 2.70,6.77 2.58,2.71 5.98,2.63 7.49,2.91 5.43,.52 23.10,.68 23.12,.68 .00,-1.3e-5 14.29,-0.02 23.81,-0.71 1.32,-0.15 4.22,-0.17 6.81,-2.89 2.03,-2.07 2.70,-6.77 2.70,-6.77 0,0 .67,-5.52 .67,-11.04 l 0,-5.17 c 0,-5.52 -0.67,-11.04 -0.67,-11.04 0,0 -0.66,-4.70 -2.70,-6.77 C 62.03,.86 59.13,.84 57.80,.69 48.28,0 34.00,0 34.00,0 33.97,0 19.69,0 10.18,.69 8.85,.84 5.95,.86 3.36,3.58 1.32,5.65 .66,10.35 .66,10.35 c 0,0 -0.55,4.50 -0.66,9.45 l 0,8.36 c .10,4.94 .66,9.45 .66,9.45 z" fill="#1f1f1e" fill-opacity="0.9">
    </path>
    <path d="m 26.96,13.67 18.37,9.62 -18.37,9.55 -0.00,-19.17 z" fill="#fff">
    </path>
    <path d="M 45.02,23.46 45.32,23.28 26.96,13.67 43.32,24.34 45.02,23.46 z" fill="#ccc">
    </path>
   </svg>
  </button>
 </div>
</div><p></p>
<p dir="auto"><code>h3c_contest_2d_r2_1344x768_12s_crf10.mp4</code> · 37,622,992 B · sha256 <code>770eaa91…003b439</code> · h264 1344×768 294 帧 24.4 Mbps + AAC 32 kHz</p>
<p dir="auto"><strong>生成视频二：3D CGI 双人对战</strong></p>
<p dir="auto"><a href="https://youtu.be/K3xQ_p3wjEk" rel="nofollow ugc"><i class="fa fa-youtube" aria-hidden="true"></i> Youtube Video</a></p><div class="js-lazyYT lazyYT-container" data-youtube-id="K3xQ_p3wjEk" data-width="640" data-height="360" data-parameters style="width:640px;padding-bottom:360px">
 <div class="ytp-thumbnail lazyYT-image-loaded" style="background-image:url(&quot;https://i.ytimg.com/vi/K3xQ_p3wjEk/hqdefault.jpg&quot;)">
  <button class="ytp-large-play-button ytp-button" tabindex="23" aria-live="assertive" style="transform:scale(0.85)" onclick="$(this).lazyYT(this);return false;">
   <svg height="100%" version="1.1" viewbox="0 0 68 48" width="100%">
    <path class="ytp-large-play-button-bg" d="m .66,37.62 c 0,0 .66,4.70 2.70,6.77 2.58,2.71 5.98,2.63 7.49,2.91 5.43,.52 23.10,.68 23.12,.68 .00,-1.3e-5 14.29,-0.02 23.81,-0.71 1.32,-0.15 4.22,-0.17 6.81,-2.89 2.03,-2.07 2.70,-6.77 2.70,-6.77 0,0 .67,-5.52 .67,-11.04 l 0,-5.17 c 0,-5.52 -0.67,-11.04 -0.67,-11.04 0,0 -0.66,-4.70 -2.70,-6.77 C 62.03,.86 59.13,.84 57.80,.69 48.28,0 34.00,0 34.00,0 33.97,0 19.69,0 10.18,.69 8.85,.84 5.95,.86 3.36,3.58 1.32,5.65 .66,10.35 .66,10.35 c 0,0 -0.55,4.50 -0.66,9.45 l 0,8.36 c .10,4.94 .66,9.45 .66,9.45 z" fill="#1f1f1e" fill-opacity="0.9">
    </path>
    <path d="m 26.96,13.67 18.37,9.62 -18.37,9.55 -0.00,-19.17 z" fill="#fff">
    </path>
    <path d="M 45.02,23.46 45.32,23.28 26.96,13.67 43.32,24.34 45.02,23.46 z" fill="#ccc">
    </path>
   </svg>
  </button>
 </div>
</div><p></p>
<p dir="auto"><code>h3c_contest_r2_1344x768_12s_crf10.mp4</code> · 36,393,959 B · sha256 <code>6721794f…6dbba84c</code> · h264 1344×768 294 帧 23.6 Mbps + AAC 32 kHz</p>
<p dir="auto">产物完整性：两个文件均含 <code>ftyp+moov+free+mdat</code>（mdat 占 99.97%），<code>ffmpeg -f null</code> 全片解码零报错。</p>
<hr />
<blockquote>
<p dir="auto">37 分钟一条 12 秒，谈不上快，但都跑在自己机器上，日志全留、哈希全留、瑕疵全写。</p>
</blockquote>
]]></description><link>https://lcz.me/topic/3074</link><generator>RSS for Node</generator><lastBuildDate>Thu, 19 Nov 2026 00:24:25 GMT</lastBuildDate><atom:link href="https://lcz.me/topic/3074.rss" rel="self" type="application/rss+xml"/><pubDate>Sat, 14 Nov 2026 16:37:11 GMT</pubDate><ttl>60</ttl><item><title><![CDATA[Reply to 【H3剑心比赛】M5 Ultra 256G H3.C引擎跑MiniMax H3模型生成2D/3D 12秒视频 on Sun, 15 Nov 2026 07:07:39 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/cc-tgo" aria-label="Profile: cc-tgo">@<bdi>cc-tgo</bdi></a></p>
<p dir="auto">sorry 我还不大了解apple 生态的困难。。。。<br />
但是我觉得他的手机生态都可以做到plug and play, 我觉得应该ai 生态也是可以。。。看以后会不会有官方给的工作流</p>
<p dir="auto">nvidia就是，安装后跑就对了</p>
]]></description><link>https://lcz.me/post/33475</link><guid isPermaLink="true">https://lcz.me/post/33475</guid><dc:creator><![CDATA[imbiplaza ASUS]]></dc:creator><pubDate>Sun, 15 Nov 2026 07:07:39 GMT</pubDate></item><item><title><![CDATA[Reply to 【H3剑心比赛】M5 Ultra 256G H3.C引擎跑MiniMax H3模型生成2D/3D 12秒视频 on Mon, 16 Nov 2026 07:07:31 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/imbiplaza-asus" aria-label="Profile: imbiplaza-ASUS">@<bdi>imbiplaza-ASUS</bdi></a> 为了参赛，用MiniMax H3 本地生成 12 秒 1344×768 视频，评估测试过三条路线：ComfyUI、MLX-Serve、h3.c 命令行。</p>
<p dir="auto">ComfyUI 被否是有硬伤：加载 H3 需常驻 100+ GB 内存，与本机 mlx-serve 的 agent 无法共存，实测 OOM 和后端崩溃，加上托管不自动拉起，我对comfyUI是完全陌生，要纯手动也就给否了，能让agent干的就不手动。</p>
<p dir="auto">MLX-Serve 日常生视频是主力，但参赛没用：它挂载的是 8bit/4bit 量化卷，参赛证据包要求官方 BF16 快照逐文件对账；且它是 HTTP 黑盒，步数日程、加速开关这些参赛要展示的细节不可暴露，最长生成单段为10.8秒，还常碰到有尾帧画面质量坍塌出马赛克色块。</p>
<p dir="auto">最终选 h3.c 原生 C/Metal 引擎跑官方 BF16 快照：内存可控、参数全暴露、证据链完整。共出两卷：第一卷双参考图模式（Ref2VA）全程 37分 ，第 2 卷双参考图全程 37 分 17 秒。为了测试去视频网格最长的一次是 BF16 无损对照卷全精度权重直载、30 步、关速度复用，从启动到落盘 1 小时 48 分钟（去噪独占 1 小时 43 分），用来验证量化与加速开关对画质的真实影响。结果没太大差别。</p>
<p dir="auto">单测试纯生成视频耗时6小时左右生成了17个视频。要实用化，我还是等生态成熟跟优化到位了再说。测试不是为了纯参赛，也是想看看这台机器在MiniMax H3模型视频生成到底可不可用。</p>
<p dir="auto">生态期望：H3 开放权重后 Mac 端已长出 h3.c、MLX-Serve、ComfyUI 节点、h3 studio 多个驱动面，方向是有了；希望官方或社区把 BF16 直跑的内存账压下来、出步数蒸馏版，把无损级出 12 秒片子从近两小时拉到几分钟到十几分钟级，让本地出片从"能跑通"变成"日常能用"。</p>
]]></description><link>https://lcz.me/post/33472</link><guid isPermaLink="true">https://lcz.me/post/33472</guid><dc:creator><![CDATA[cc tgo]]></dc:creator><pubDate>Mon, 16 Nov 2026 07:07:31 GMT</pubDate></item><item><title><![CDATA[Reply to 【H3剑心比赛】M5 Ultra 256G H3.C引擎跑MiniMax H3模型生成2D/3D 12秒视频 on Sun, 15 Nov 2026 05:24:30 GMT]]></title><description><![CDATA[<p dir="auto">尽量使用坊间的优化，b16不代表就是好，尤其是用了b16也出现劣质的网格，已经消耗了那么多的时间，还会出现画面眼见得瑕疵，其实是不合格的，我给一个提点：</p>
<p dir="auto">不要出现肉眼可以见得网格这是我的起码最低标准</p>
<p dir="auto">2d<br />
<img src="https://upload.lcz.me/uploads/6f69ba8f-11d2-4706-a520-13078e87f763.jpeg" alt="image.jpeg" class="img-markdown  img-fluid" /></p>
<p dir="auto">3d<br />
<img src="https://upload.lcz.me/uploads/9e069971-b4b9-4311-9c03-9da6fe2ea528.jpeg" alt="image.jpeg" class="img-markdown  img-fluid" /></p>
]]></description><link>https://lcz.me/post/33459</link><guid isPermaLink="true">https://lcz.me/post/33459</guid><dc:creator><![CDATA[imbiplaza ASUS]]></dc:creator><pubDate>Sun, 15 Nov 2026 05:24:30 GMT</pubDate></item><item><title><![CDATA[Reply to 【H3剑心比赛】M5 Ultra 256G H3.C引擎跑MiniMax H3模型生成2D/3D 12秒视频 on Sun, 15 Nov 2026 05:09:35 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/ralfzhu" aria-label="Profile: ralfzhu">@<bdi>ralfzhu</bdi></a> 要不看时间的话，权重更大效果是会更好。</p>
]]></description><link>https://lcz.me/post/33457</link><guid isPermaLink="true">https://lcz.me/post/33457</guid><dc:creator><![CDATA[terry]]></dc:creator><pubDate>Sun, 15 Nov 2026 05:09:35 GMT</pubDate></item><item><title><![CDATA[Reply to 【H3剑心比赛】M5 Ultra 256G H3.C引擎跑MiniMax H3模型生成2D/3D 12秒视频 on Sat, 14 Nov 2026 23:27:10 GMT]]></title><description><![CDATA[<p dir="auto">这个真不错！是不是模型权重比我们这种简化版的要大一个量级的关系哦？</p>
]]></description><link>https://lcz.me/post/33377</link><guid isPermaLink="true">https://lcz.me/post/33377</guid><dc:creator><![CDATA[ralfzhu]]></dc:creator><pubDate>Sat, 14 Nov 2026 23:27:10 GMT</pubDate></item><item><title><![CDATA[Reply to 【H3剑心比赛】M5 Ultra 256G H3.C引擎跑MiniMax H3模型生成2D/3D 12秒视频 on Sat, 14 Nov 2026 23:12:09 GMT]]></title><description><![CDATA[<p dir="auto">通过初审，正式参赛，建议20积分，比赛结束后统一结算。</p>
]]></description><link>https://lcz.me/post/33372</link><guid isPermaLink="true">https://lcz.me/post/33372</guid><dc:creator><![CDATA[terry]]></dc:creator><pubDate>Sat, 14 Nov 2026 23:12:09 GMT</pubDate></item><item><title><![CDATA[Reply to 【H3剑心比赛】M5 Ultra 256G H3.C引擎跑MiniMax H3模型生成2D/3D 12秒视频 on Sat, 14 Nov 2026 23:11:39 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 e3074-3451-1</p>
</blockquote>
]]></description><link>https://lcz.me/post/33371</link><guid isPermaLink="true">https://lcz.me/post/33371</guid><dc:creator><![CDATA[Xiaote]]></dc:creator><pubDate>Sat, 14 Nov 2026 23:11:39 GMT</pubDate></item><item><title><![CDATA[Reply to 【H3剑心比赛】M5 Ultra 256G H3.C引擎跑MiniMax H3模型生成2D/3D 12秒视频 on Sat, 14 Nov 2026 17:55:16 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> 已改重发</p>
]]></description><link>https://lcz.me/post/33363</link><guid isPermaLink="true">https://lcz.me/post/33363</guid><dc:creator><![CDATA[cc tgo]]></dc:creator><pubDate>Sat, 14 Nov 2026 17:55:16 GMT</pubDate></item><item><title><![CDATA[Reply to 【H3剑心比赛】M5 Ultra 256G H3.C引擎跑MiniMax H3模型生成2D/3D 12秒视频 on Sat, 14 Nov 2026 16:50:31 GMT]]></title><description><![CDATA[<p dir="auto">作品2很好，可以通过，但是作品1需要重发，严格按照要求来，使用后剑心和志志雄。另外附上json工作流。</p>
]]></description><link>https://lcz.me/post/33343</link><guid isPermaLink="true">https://lcz.me/post/33343</guid><dc:creator><![CDATA[terry]]></dc:creator><pubDate>Sat, 14 Nov 2026 16:50:31 GMT</pubDate></item></channel></rss>