<blockquote>
⚡ Compact <code>~345 MB</code> 3–4 Step All-in-One Acceleration + De-Quantization Suite for MiniMax H3 (33B Video + 32kHz Stereo Audio DiT)! 
Includes both dedicated editions as version tabs above (plus Ready-to-Run ComfyUI 0.33.0+ Workflows attached under <code>Files</code>): 
• Tab 1 — <code>FL2VA / I2VA / T2VA</code> Edition (<code>346.7 MB</code>): Pure Text-to-Video, Single-Frame Image-to-Video, and First+Last Frame up to 15 Seconds (<code>362f @ 24fps</code> + Stereo Dialogue Audio). 
• Tab 2 — <code>Ref2VA</code> Multi-Ref &amp; Lip-Sync Edition (<code>340.3 MB</code>): Multi-Subject Character Turnaround + Scene Reference + Audio Lip-Sync up to 15 Seconds. 
🔥 Universal Base Precision Compatibility: Trained with online quantization-aware perturbation across <code>Pure BF16</code> + <code>Block-256 Hadamard ConvRot INT8</code> + <code>GGUF Q4_K_M</code> — one LoRA works seamlessly across <code>&lt; 9 GB VRAM</code> consumer GPUs (<code>UnetLoaderGGUFAdvanced</code>) and full <code>BF16 / INT8</code> setups! 
🇨🇳 中文速览：NineCrowns (<code>[CROWN]</code>) 出品的 MiniMax H3 3–4 步全合一超清加速 + 去量化伪影 LoRA 套件（仅约 <code>345 MB</code>，保留 <code>99.96%+</code> 谱能量）！上方双版本标签分别对应 <code>FL2VA / 首尾帧 / 单图I2V / 纯文生T2V</code> 与 <code>Ref2VA 多图主体参考 / 对口型</code>，完美通吃 <code>BF16</code>、<code>ConvRot INT8</code> 与 <code>GGUF Q4_K_M</code>（显存占用 &lt; 9GB） 三种精度底模，附带 ComfyUI 0.33.0+ 开箱即用工作流压缩包！
</blockquote>

<h3>🚀 Why CROWN HyperPDMD-HD? (Solving 3–4 Step Collapse &amp; Q4 Quantization Blur)</h3>
<ol>
 <li>True 3–4 Step Trajectory &amp; Mid-Clip Sharpness up to 15 Seconds (<code>362f @ 24fps</code>): Standard 3–4 step MiniMax H3 without LoRA freezes on single-frame I2V, melts into faceless blobs on pure T2V, and suffers severe ghosting/blur in the middle 5s–10s of 15-second FL2V clips (<code>Laplacian = 22 ~ 48</code>). CROWN HyperPDMD-HD delivers <code>6.5× ~ 10×</code> higher optical sharpness with full dynamic motion and synchronized <code>32kHz</code> stereo dialogue audio in just 3–4 Euler steps (<code>CFG = 1.0</code>).</li>
 <li>Online Multi-Precision De-Quantization Training (<code>BF16 + ConvRot INT8 + GGUF Q4_K_M</code>): Distilled on <code>8×NVIDIA B200</code> clusters using Projected Distribution Matching Distillation (<code>HyperPDMD</code>) + 5-frame 3D-Conv VideoVAE spectral discriminator while dynamically perturbing student base weights across <code>45% GGUF Q4_K_M</code>, <code>35% Block-256 Hadamard ConvRot INT8</code>, and <code>20% Pure BF16</code>. Eliminates 4-bit quantization haze while adding <code>0.00 GB</code> extra VRAM when loaded with <code>UnetLoaderGGUFAdvanced (patch_dtype="target")</code>.</li>
 <li>Compact Energy-Optimal SVD Rank Compression (<code>~345 MB</code> vs <code>2.0–3.9 GB</code> Bloated LoRAs): Compressed from <code>Rank-64</code> Cosine-Tail Checkpoint Soup via exact FP64 QR-SVD to <code>avg_rank = 38.0</code> across all 208 DiT + TokenRefiner layers (<code>comfy.lora.load_lora = 208/208 = 100%</code>), retaining <code>99.96%–99.98%</code> Frobenius energy at 1/10th the file size.</li>
</ol>

<hr />
🛠️ Recommended ComfyUI Settings / 推荐推理参数
<ul>
 <li>Base Models Supported (全兼容底模):
 <ul>
 <li>Consumer <code>&lt; 9 GB VRAM</code>: <code>minimax-h3-{fl2va|ref2va}-pruned-Q4_K_M.gguf</code> via <code>UnetLoaderGGUFAdvanced</code> (set <code>patch_dtype = "target"</code>)</li>
 <li>16–24 GB VRAM: <code>minimax_h3_{fl2va|ref2va}_pruned_int8_convrot.safetensors</code></li>
 <li>Workstation / Cloud: <code>minimax_h3_{fl2va|ref2va}_pruned_bf16.safetensors</code></li>
 </ul>
 </li>
 <li>LoRA Strength (推荐权重): <code>1.0</code> (<code>0.95 ~ 1.05</code>)</li>
 <li>Sampler &amp; Scheduler (采样器与调度器): <code>SamplerCustomAdvanced</code> + <code>BasicGuider (CFG = 1.0)</code> + <code>KSamplerSelect ("euler")</code> + <code>BetaSamplingScheduler (steps = 4 or 3, alpha = 0.6, beta = 0.6)</code></li>
 <li><code>MiniMaxH3SigmaShift</code>:
 <ul>
 <li><code>FL2VA / I2VA / T2VA</code>: <code>shift_video = 12.0, shift_audio = 3.0</code></li>
 <li><code>Ref2VA</code> (768P HD Multi-Ref): <code>shift_video = 9.0 ~ 12.0, shift_audio = 3.0</code>, and set <code>ref_image_size = "max"</code> on <code>MiniMaxH3ReferenceToVideo</code> for 2K pore-level facial fidelity!</li>
 </ul>
 </li>
</ul>

CROWN HyperPDMD-HD 3-4 Step All-in-One Turbo & De-Quantization LoRA | MiniMax H3 (BF16 / INT8 / GGUF Q4)

[EN] v1.0 <code>Ref2VA</code> Multi-Subject Reference &amp; Audio Lip-Sync 3–4 Step Edition (<code>340.3 MB</code>, <code>avg_rank=38.0</code> retaining <code>99.98%</code> Frobenius energy of 2,000-step Rank-64 Soup). Compatible with <code>BF16</code>, <code>ConvRot INT8</code>, and <code>GGUF Q4_K_M</code>. Set <code>ref_image_size='max'</code> + <code>shift_video=9.0</code> for razor-sharp 768P multi-character fidelity. [中文] v1.0 <code>Ref2VA 多图主体参考与音频对口型</code> 3–4 步专版（<code>340.3 MB</code>，保留 2000 步母模 <code>99.98%</code> 能量），全面兼容 <code>BF16 / INT8 / GGUF Q4_K_M</code>，建议配合 <code>ref_image_size='max'</code> + <code>shift_video=9.0</code> 获得毛孔级超清人像还原。

Files with SHA256: dac1d81d1a60631be9c04ab010b5c844644c53a97c4b1dbe408d867d6501e45d