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

作者：NineCrowns

⚡ Compact ~345 MB 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 Files):

• Tab 1 — FL2VA / I2VA / T2VA Edition (346.7 MB): Pure Text-to-Video, Single-Frame Image-to-Video, and First+Last Frame up to 15 Seconds (362f @ 24fps + Stereo Dialogue Audio).

• Tab 2 — Ref2VA Multi-Ref & Lip-Sync Edition (340.3 MB): Multi-Subject Character Turnaround + Scene Reference + Audio Lip-Sync up to 15 Seconds.

🔥 Universal Base Precision Compatibility: Trained with online quantization-aware perturbation across Pure BF16 + Block-256 Hadamard ConvRot INT8 + GGUF Q4_K_M — one LoRA works seamlessly across < 9 GB VRAM consumer GPUs (UnetLoaderGGUFAdvanced) and full BF16 / INT8 setups!

🇨🇳 中文速览：NineCrowns ([CROWN]) 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 345 MB，保留 99.96%+ 谱能量）！上方双版本标签分别对应 FL2VA / 首尾帧 / 单图I2V / 纯文生T2V 与 Ref2VA 多图主体参考 / 对口型，完美通吃 BF16、ConvRot INT8 与 GGUF Q4_K_M（显存占用 < 9GB） 三种精度底模，附带 ComfyUI 0.33.0+ 开箱即用工作流压缩包！

🚀 Why CROWN HyperPDMD-HD? (Solving 3–4 Step Collapse & Q4 Quantization Blur)

• True 3–4 Step Trajectory & Mid-Clip Sharpness up to 15 Seconds (362f @ 24fps): 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 (Laplacian = 22 ~ 48). CROWN HyperPDMD-HD delivers 6.5× ~ 10× higher optical sharpness with full dynamic motion and synchronized 32kHz stereo dialogue audio in just 3–4 Euler steps (CFG = 1.0).

• Online Multi-Precision De-Quantization Training (BF16 + ConvRot INT8 + GGUF Q4_K_M): Distilled on 8×NVIDIA B200 clusters using Projected Distribution Matching Distillation (HyperPDMD) + 5-frame 3D-Conv VideoVAE spectral discriminator while dynamically perturbing student base weights across 45% GGUF Q4_K_M, 35% Block-256 Hadamard ConvRot INT8, and 20% Pure BF16. Eliminates 4-bit quantization haze while adding 0.00 GB extra VRAM when loaded with UnetLoaderGGUFAdvanced (patch_dtype="target").

• Compact Energy-Optimal SVD Rank Compression (~345 MB vs 2.0–3.9 GB Bloated LoRAs): Compressed from Rank-64 Cosine-Tail Checkpoint Soup via exact FP64 QR-SVD to avg_rank = 38.0 across all 208 DiT + TokenRefiner layers (comfy.lora.load_lora = 208/208 = 100%), retaining 99.96%–99.98% Frobenius energy at 1/10th the file size.

🛠️ Recommended ComfyUI Settings / 推荐推理参数

• Base Models Supported (全兼容底模):

• Consumer < 9 GB VRAM: minimax-h3-{fl2va|ref2va}-pruned-Q4_K_M.gguf via UnetLoaderGGUFAdvanced (set patch_dtype = "target")

• 16–24 GB VRAM: minimax_h3_{fl2va|ref2va}_pruned_int8_convrot.safetensors

• Workstation / Cloud: minimax_h3_{fl2va|ref2va}_pruned_bf16.safetensors

• LoRA Strength (推荐权重): 1.0 (0.95 ~ 1.05)

• Sampler & Scheduler (采样器与调度器): SamplerCustomAdvanced + BasicGuider (CFG = 1.0) + KSamplerSelect ("euler") + BetaSamplingScheduler (steps = 4 or 3, alpha = 0.6, beta = 0.6)

• MiniMaxH3SigmaShift:

• FL2VA / I2VA / T2VA: shift_video = 12.0, shift_audio = 3.0

• Ref2VA (768P HD Multi-Ref): shift_video = 9.0 ~ 12.0, shift_audio = 3.0, and set ref_image_size = "max" on MiniMaxH3ReferenceToVideo for 2K pore-level facial fidelity!