<?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 5080 16GB 单卡｜MiniMax H3 3D / 2D｜灰盒深度控动作＋STEP600 8 步]]></title><description><![CDATA[<p dir="auto">3D CGI 成片：<a href="https://youtu.be/lXGIlkPRJN4" rel="nofollow ugc">https://youtu.be/lXGIlkPRJN4</a><br />
传统 2D 动画成片：<a href="https://youtu.be/GaWO4hwy53A" rel="nofollow ugc">https://youtu.be/GaWO4hwy53A</a></p>
<blockquote>
<p dir="auto">成片 = H3 原样画面（1344×768，取前 288 帧＝12.00 秒）＋后期配音、字幕、RTX 超分到 2560×1440。画面本身没有补帧、裁切、重绘。和别人不一样的地方在下面「说明」里照实写了。</p>
</blockquote>
<p dir="auto"><strong>关键帧（3D）</strong><br />
<img src="https://upload.lcz.me/uploads/7dfba7f3-ad1b-45b5-b2ab-8715f24fb2e2.jpg" alt="关键帧_3D.jpg" class="img-markdown  img-fluid" /></p>
<p dir="auto"><strong>关键帧（2D）</strong><br />
<img src="https://upload.lcz.me/uploads/b652ff0b-1292-4415-8315-2ec57cc5d273.jpg" alt="关键帧_2D.jpg" class="img-markdown  img-fluid" /></p>
<h2>硬件</h2>
<ul>
<li>显卡：RTX 5080 16GB（驱动 616.92，报告 16,303 MiB）</li>
<li>CPU：Ryzen 9 9900X｜内存：128GB｜SSD：Crucial P310 2TB</li>
<li>系统：Windows 11 Pro（build 26300）｜ComfyUI 0.44.0 便携版｜torch 2.14.1+cu130｜Python 3.13</li>
</ul>
<h2>工作流（两条完全相同，只换画风和参考图）</h2>
<ul>
<li>底模：<code>minimax_h3_hybrid_fl2va_ref2va_b25-49-int8</code></li>
<li>加速 LoRA：larryvrh <code>minimax_h3_turbo_v4_step600_ema</code>，强度 1.0（节点 <code>MiniMaxH3TurboLoRA</code>＋<code>MiniMaxH3TurboSampler</code>，来自 ComfyUI-MiniMax-H3-Turbo）</li>
<li>采样：8 步，simple，BasicGuider，种子 <code>1</code>（两条同一个）</li>
<li>文本编码器 <code>qwen3vl_32b_minimax_h3_nvfp4_awq</code>；视频 VAE <code>fp16</code>＋VAEDecodeTiled；音频 VAE fp32</li>
<li>参考图：ComfyUI 自带 <code>MiniMax H3 Reference to Video</code>
<ul>
<li>3D：剑心手办＋志志雄 3 张手办（共 4 张）</li>
<li>2D：剑心 2 张动画图＋志志雄 3 张手办（共 5 张）</li>
</ul>
</li>
<li>控制：Fun ControlNet Union 2.0（剪枝 int8），<strong>自己做的灰盒深度视频</strong>，强度 <code>0.6</code></li>
<li>1344×768，生成 294 帧（H3 要 17n+5 帧），导出前 288 帧＝24fps 12.00 秒</li>
<li>提示词：比赛原文逐字不改，只把 Style 二选一</li>
<li>工作流 JSON（API 格式）在文末</li>
</ul>
<h2>加速手段</h2>
<ul>
<li>int8 剪枝底模＋STEP600 加速 LoRA 8 步</li>
<li>comfy kitchen attention（全量注意力，不用稀疏）</li>
<li>16G 显存跑 1344×768×294 帧＋控制网：<code>--disable-dynamic-vram --reserve-vram 9</code>，再加两个分块补丁（MLP 和 LoRA 的线性层按 16384 个 token 分块算，数学完全一样，只是省显存）</li>
</ul>
<h2>数据</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>版本</th>
<th style="text-align:right">Total Time</th>
<th style="text-align:right">ComfyUI 执行</th>
<th style="text-align:right">8 步采样</th>
<th style="text-align:right">显存峰值</th>
<th style="text-align:right">内存峰值</th>
<th style="text-align:right">显卡功耗</th>
<th style="text-align:right">最高温度</th>
</tr>
</thead>
<tbody>
<tr>
<td>3D</td>
<td style="text-align:right">762.4 s</td>
<td style="text-align:right">12:39</td>
<td style="text-align:right">10:17（77.2 s/步）</td>
<td style="text-align:right">15,803 MiB</td>
<td style="text-align:right">74.3 GB</td>
<td style="text-align:right">344 W</td>
<td style="text-align:right">79°C</td>
</tr>
<tr>
<td>2D</td>
<td style="text-align:right">719.8 s</td>
<td style="text-align:right">11:57</td>
<td style="text-align:right">10:23（78.0 s/步）</td>
<td style="text-align:right">15,779 MiB</td>
<td style="text-align:right">69.0 GB</td>
<td style="text-align:right">346 W</td>
<td style="text-align:right">79°C</td>
</tr>
</tbody>
</table>
<p dir="auto">Total Time＝提交工作流到 MP4 写完（含读模型、文本和参考图编码、采样、视频和声音解码）。3D 是 ComfyUI 启动后第一条，所以多了读模型的时间。</p>
<p dir="auto"><strong>运行日志（3D）</strong><br />
<img src="https://upload.lcz.me/uploads/fed5b010-7920-4b29-952b-f105af52fbce.png" alt="日志_3D.png" class="img-markdown  img-fluid" /></p>
<p dir="auto"><strong>运行日志（2D）</strong><br />
<img src="https://upload.lcz.me/uploads/c4b95d1a-a441-43ef-a3e0-3d673d8cd3fc.png" alt="日志_2D.png" class="img-markdown  img-fluid" /></p>
<h2>说明：和别人不一样的地方（照实写）</h2>
<ol>
<li><strong>用了灰盒深度控制。</strong> 自己用 Python 画了方块人灰盒（红＝剑心、紫＝志志雄），排好 12 秒的走位和机位，导出深度视频喂给控制网。强度 1.0 时两个人会认反（深度图没颜色），降到 0.6 才对。提示词没改。</li>
</ol>
<p dir="auto"><img src="https://upload.lcz.me/uploads/32b04245-3b6b-48bd-99bd-906c9fc3d991.jpg" alt="灰盒预览.jpg" class="img-markdown  img-fluid" /><br />
2. <strong>声音全是后期配的。</strong> H3 自带的声音不用；日语喊声用 Gemini TTS，撞击声用 Stable Audio 3 Medium 做，再按画面逐帧对上撞剑时刻；中文字幕。<br />
3. <strong>放大。</strong> 成片用 NVIDIA RTX 超分（ComfyUI 的 <code>RTXVideoSuperResolution</code> 节点）放大到 2560×1440，加了一点对比和锐化。</p>
<h2>过程：快动作「划痕」是怎么解决的</h2>
<p dir="auto">一开始用 <code>minimax_h3_ref2v_turbo_4step_v0.1</code> 走 4 步，快动作时人物身上一道道划痕，像油画笔触（发过求助帖 <a href="https://lcz.me/topic/3092">https://lcz.me/topic/3092</a> ）。<br />
按 <a class="plugin-mentions-user plugin-mentions-a" href="/user/perter" aria-label="Profile: perter">@<bdi>perter</bdi></a> 的建议换成 larryvrh 的 STEP600、走 8 步，划痕基本没了。同一个 4 步 LoRA 加到 8 步几乎没用，问题出在那个 LoRA 本身。谢谢 perter！<br />
左＝旧版（4 步 LoRA），右＝这次（STEP600 8 步）：<br />
<img src="https://upload.lcz.me/uploads/200b2b7e-0b55-47a3-b38c-310e01cb4dd3.jpg" alt="旧版vsSTEP600_3D.jpg" class="img-markdown  img-fluid" /></p>
<h2>自评</h2>
<ul>
<li>好的：动作顺序按灰盒走（冲刺 → 横扫 → 相撞 → 对砍 → 落地收势），快动作清楚了；3D 手办质感和火光比较足。</li>
<li>不够的：2D 版剑心头发在 3～8.5 秒偏金黄；2D 有两处灰色冲击波环；3D 撞剑白光被背景的火盖住、不够亮；加了灰盒后每条要 12 分钟左右，比不加慢。</li>
</ul>
<h2>复现</h2>
<ul>
<li>参考图：比赛原帖 6 张（3D 用 4 张、2D 用 5 张，见上）</li>
<li>灰盒深度视频原件 2.6MB，超过论坛 2MB 上限；需要的话私信我</li>
<li>工作流 JSON（3D，API 格式）：</li>
</ul>
<pre><code class="language-json">{
 "unet": {
  "class_type": "UNETLoader",
  "inputs": {
   "unet_name": "minimax_h3_hybrid_fl2va_ref2va_b25-49-int8.safetensors",
   "weight_dtype": "default"
  }
 },
 "clip": {
  "class_type": "CLIPLoader",
  "inputs": {
   "clip_name": "qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors",
   "type": "minimax",
   "device": "default"
  }
 },
 "vae": {
  "class_type": "VAELoader",
  "inputs": {
   "vae_name": "minimax_h3_video_vae_fp16.safetensors"
  }
 },
 "cond": {
  "class_type": "MiniMaxH3ReferenceToVideo",
  "inputs": {
   "clip": [
    "clip",
    0
   ],
   "vae": [
    "vae",
    0
   ],
   "audio_vae": [
    "avae",
    0
   ],
   "prompt": "(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.",
   "width": 1344,
   "height": 768,
   "length": 294,
   "ref_image_size": "match",
   "ref_images.ref_image_0": [
    "rimg0",
    0
   ],
   "ref_images.ref_image_1": [
    "rimg1",
    0
   ],
   "ref_images.ref_image_2": [
    "rimg2",
    0
   ],
   "ref_images.ref_image_3": [
    "rimg3",
    0
   ]
  }
 },
 "noise": {
  "class_type": "RandomNoise",
  "inputs": {
   "noise_seed": 1
  }
 },
 "dec": {
  "class_type": "VAEDecodeTiled",
  "inputs": {
   "samples": [
    "ksa",
    0
   ],
   "vae": [
    "vae",
    0
   ],
   "tile_size": 512,
   "overlap": 64,
   "temporal_size": 64,
   "temporal_overlap": 8
  }
 },
 "save": {
  "class_type": "SaveVideo",
  "inputs": {
   "video": [
    "mk",
    0
   ],
   "filename_prefix": "h3test/K3D_s600",
   "format": "auto",
   "codec": "auto"
  }
 },
 "lora": {
  "class_type": "MiniMaxH3TurboLoRA",
  "inputs": {
   "model": [
    "unet",
    0
   ],
   "lora_name": "minimax_h3_turbo_v4_step600_ema.safetensors",
   "strength": 1.0,
   "low_vram": false
  }
 },
 "attention": {
  "class_type": "ModelAttentionBackend",
  "inputs": {
   "model": [
    "lora",
    0
   ],
   "attention": "comfy kitchen attention"
  }
 },
 "sampler": {
  "class_type": "MiniMaxH3TurboSampler",
  "inputs": {}
 },
 "patch": {
  "class_type": "ModelPatchLoader",
  "inputs": {
   "name": "minimax_h3_fun_controlnet_union_2.0_pruned_int8_convrot.safetensors"
  }
 },
 "lv": {
  "class_type": "LoadVideo",
  "inputs": {
   "file": "kenshin_ctrl_v3b_depth_1344x768_294f.mp4"
  }
 },
 "comp": {
  "class_type": "GetVideoComponents",
  "inputs": {
   "video": [
    "lv",
    0
   ]
  }
 },
 "cn": {
  "class_type": "MiniMaxH3FunControlNetApply",
  "inputs": {
   "model": [
    "attention",
    0
   ],
   "model_patch": [
    "patch",
    0
   ],
   "vae": [
    "vae",
    0
   ],
   "strength": 0.6,
   "start_percent": 0.0,
   "end_percent": 1.0,
   "control_video": [
    "comp",
    0
   ]
  }
 },
 "avae": {
  "class_type": "VAELoader",
  "inputs": {
   "vae_name": "minimax_h3_audio_vae_fp32.safetensors"
  }
 },
 "rimg0": {
  "class_type": "LoadImage",
  "inputs": {
   "image": "kenshin_ref3.jpg"
  }
 },
 "rimg1": {
  "class_type": "LoadImage",
  "inputs": {
   "image": "kenshin_ref4.jpg"
  }
 },
 "rimg2": {
  "class_type": "LoadImage",
  "inputs": {
   "image": "kenshin_ref5.jpg"
  }
 },
 "rimg3": {
  "class_type": "LoadImage",
  "inputs": {
   "image": "kenshin_ref6.jpg"
  }
 },
 "sched": {
  "class_type": "BasicScheduler",
  "inputs": {
   "model": [
    "cn",
    0
   ],
   "scheduler": "simple",
   "steps": 8,
   "denoise": 1.0
  }
 },
 "guider": {
  "class_type": "BasicGuider",
  "inputs": {
   "model": [
    "cn",
    0
   ],
   "conditioning": [
    "cond",
    0
   ]
  }
 },
 "ksa": {
  "class_type": "SamplerCustomAdvanced",
  "inputs": {
   "noise": [
    "noise",
    0
   ],
   "guider": [
    "guider",
    0
   ],
   "sampler": [
    "sampler",
    0
   ],
   "sigmas": [
    "sched",
    0
   ],
   "latent_image": [
    "cond",
    1
   ]
  }
 },
 "adec": {
  "class_type": "VAEDecodeAudio",
  "inputs": {
   "samples": [
    "ksa",
    0
   ],
   "vae": [
    "avae",
    0
   ]
  }
 },
 "mk": {
  "class_type": "CreateVideo",
  "inputs": {
   "images": [
    "dec",
    0
   ],
   "fps": 24.0,
   "audio": [
    "adec",
    0
   ]
  }
 }
}
</code></pre>
<ul>
<li>工作流 JSON（2D，API 格式）：</li>
</ul>
<pre><code class="language-json">{
 "unet": {
  "class_type": "UNETLoader",
  "inputs": {
   "unet_name": "minimax_h3_hybrid_fl2va_ref2va_b25-49-int8.safetensors",
   "weight_dtype": "default"
  }
 },
 "clip": {
  "class_type": "CLIPLoader",
  "inputs": {
   "clip_name": "qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors",
   "type": "minimax",
   "device": "default"
  }
 },
 "vae": {
  "class_type": "VAELoader",
  "inputs": {
   "vae_name": "minimax_h3_video_vae_fp16.safetensors"
  }
 },
 "cond": {
  "class_type": "MiniMaxH3ReferenceToVideo",
  "inputs": {
   "clip": [
    "clip",
    0
   ],
   "vae": [
    "vae",
    0
   ],
   "audio_vae": [
    "avae",
    0
   ],
   "prompt": "(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.",
   "width": 1344,
   "height": 768,
   "length": 294,
   "ref_image_size": "match",
   "ref_images.ref_image_0": [
    "rimg0",
    0
   ],
   "ref_images.ref_image_1": [
    "rimg1",
    0
   ],
   "ref_images.ref_image_2": [
    "rimg2",
    0
   ],
   "ref_images.ref_image_3": [
    "rimg3",
    0
   ],
   "ref_images.ref_image_4": [
    "rimg4",
    0
   ]
  }
 },
 "noise": {
  "class_type": "RandomNoise",
  "inputs": {
   "noise_seed": 1
  }
 },
 "dec": {
  "class_type": "VAEDecodeTiled",
  "inputs": {
   "samples": [
    "ksa",
    0
   ],
   "vae": [
    "vae",
    0
   ],
   "tile_size": 512,
   "overlap": 64,
   "temporal_size": 64,
   "temporal_overlap": 8
  }
 },
 "save": {
  "class_type": "SaveVideo",
  "inputs": {
   "video": [
    "mk",
    0
   ],
   "filename_prefix": "h3test/K2D_s600",
   "format": "auto",
   "codec": "auto"
  }
 },
 "lora": {
  "class_type": "MiniMaxH3TurboLoRA",
  "inputs": {
   "model": [
    "unet",
    0
   ],
   "lora_name": "minimax_h3_turbo_v4_step600_ema.safetensors",
   "strength": 1.0,
   "low_vram": false
  }
 },
 "attention": {
  "class_type": "ModelAttentionBackend",
  "inputs": {
   "model": [
    "lora",
    0
   ],
   "attention": "comfy kitchen attention"
  }
 },
 "sampler": {
  "class_type": "MiniMaxH3TurboSampler",
  "inputs": {}
 },
 "patch": {
  "class_type": "ModelPatchLoader",
  "inputs": {
   "name": "minimax_h3_fun_controlnet_union_2.0_pruned_int8_convrot.safetensors"
  }
 },
 "lv": {
  "class_type": "LoadVideo",
  "inputs": {
   "file": "kenshin_ctrl_v3b_depth_1344x768_294f.mp4"
  }
 },
 "comp": {
  "class_type": "GetVideoComponents",
  "inputs": {
   "video": [
    "lv",
    0
   ]
  }
 },
 "cn": {
  "class_type": "MiniMaxH3FunControlNetApply",
  "inputs": {
   "model": [
    "attention",
    0
   ],
   "model_patch": [
    "patch",
    0
   ],
   "vae": [
    "vae",
    0
   ],
   "strength": 0.6,
   "start_percent": 0.0,
   "end_percent": 1.0,
   "control_video": [
    "comp",
    0
   ]
  }
 },
 "avae": {
  "class_type": "VAELoader",
  "inputs": {
   "vae_name": "minimax_h3_audio_vae_fp32.safetensors"
  }
 },
 "rimg0": {
  "class_type": "LoadImage",
  "inputs": {
   "image": "kenshin_ref1.jpg"
  }
 },
 "rimg1": {
  "class_type": "LoadImage",
  "inputs": {
   "image": "kenshin_ref2.jpg"
  }
 },
 "rimg2": {
  "class_type": "LoadImage",
  "inputs": {
   "image": "kenshin_ref4.jpg"
  }
 },
 "rimg3": {
  "class_type": "LoadImage",
  "inputs": {
   "image": "kenshin_ref5.jpg"
  }
 },
 "rimg4": {
  "class_type": "LoadImage",
  "inputs": {
   "image": "kenshin_ref6.jpg"
  }
 },
 "sched": {
  "class_type": "BasicScheduler",
  "inputs": {
   "model": [
    "cn",
    0
   ],
   "scheduler": "simple",
   "steps": 8,
   "denoise": 1.0
  }
 },
 "guider": {
  "class_type": "BasicGuider",
  "inputs": {
   "model": [
    "cn",
    0
   ],
   "conditioning": [
    "cond",
    0
   ]
  }
 },
 "ksa": {
  "class_type": "SamplerCustomAdvanced",
  "inputs": {
   "noise": [
    "noise",
    0
   ],
   "guider": [
    "guider",
    0
   ],
   "sampler": [
    "sampler",
    0
   ],
   "sigmas": [
    "sched",
    0
   ],
   "latent_image": [
    "cond",
    1
   ]
  }
 },
 "adec": {
  "class_type": "VAEDecodeAudio",
  "inputs": {
   "samples": [
    "ksa",
    0
   ],
   "vae": [
    "avae",
    0
   ]
  }
 },
 "mk": {
  "class_type": "CreateVideo",
  "inputs": {
   "images": [
    "dec",
    0
   ],
   "fps": 24.0,
   "audio": [
    "adec",
    0
   ]
  }
 }
}
</code></pre>
]]></description><link>https://lcz.me/topic/3104</link><generator>RSS for Node</generator><lastBuildDate>Thu, 19 Nov 2026 00:16:37 GMT</lastBuildDate><atom:link href="https://lcz.me/topic/3104.rss" rel="self" type="application/rss+xml"/><pubDate>Sun, 15 Nov 2026 23:00:37 GMT</pubDate><ttl>60</ttl><item><title><![CDATA[Reply to [H3剑心比赛] RTX 5080 16GB 单卡｜MiniMax H3 3D / 2D｜灰盒深度控动作＋STEP600 8 步 on Mon, 16 Nov 2026 01:10:13 GMT]]></title><description><![CDATA[<p dir="auto">过审，20积分比赛结束后统一结算。画面不错，你的声音约束下，不过也能接受。</p>
]]></description><link>https://lcz.me/post/33646</link><guid isPermaLink="true">https://lcz.me/post/33646</guid><dc:creator><![CDATA[terry]]></dc:creator><pubDate>Mon, 16 Nov 2026 01:10:13 GMT</pubDate></item><item><title><![CDATA[Reply to [H3剑心比赛] RTX 5080 16GB 单卡｜MiniMax H3 3D / 2D｜灰盒深度控动作＋STEP600 8 步 on Mon, 16 Nov 2026 01:09:44 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 e3104-4790-1</p>
</blockquote>
]]></description><link>https://lcz.me/post/33645</link><guid isPermaLink="true">https://lcz.me/post/33645</guid><dc:creator><![CDATA[Xiaote]]></dc:creator><pubDate>Mon, 16 Nov 2026 01:09:44 GMT</pubDate></item></channel></rss>