⚑ CROWN MiniMax-H3 HyperPDMD-HD (3–4 Step All-in-One Turbo & De-Quantization LoRA Suite)

By NineCrowns ([CROWN]) β€” Compact (~345 MiB) 3–4 Step Acceleration + Nyquist-HD Sharpness + Multi-Precision (BF16 / ConvRot INT8 / GGUF Q4_K_M) De-Quantization LoRA Suite for MiniMax-H3 (33B Joint Video + 32kHz Stereo Audio DiT).


πŸ“¦ Included Weights & Workflows

File Edition / Scope Size Rank Energy Retained
CROWN_MiniMax_H3_HyperPDMD_HD_FL2VA_3_4Step_r38.safetensors FL2VA / I2VA / T2VA (Recommended Main) 346.7 MiB avg_rank=38 99.96%
CROWN_MiniMax_H3_HyperPDMD_HD_Ref2VA_3_4Step_r38.safetensors Ref2VA Multi-Ref & Lip-Sync (Recommended Main) 340.3 MiB avg_rank=38 99.98%
full_rank64/CROWN_MiniMax_H3_HyperPDMD_HD_FL2VA_3_4Step_r64_full.safetensors FL2VA / I2VA / T2VA Uncompressed Master 591.6 MiB Rank 64 100.00%
full_rank64/CROWN_MiniMax_H3_HyperPDMD_HD_Ref2VA_3_4Step_r64_full.safetensors Ref2VA Uncompressed Master 591.6 MiB Rank 64 100.00%
workflows/*.json 6 Ready-to-Run ComfyUI 0.33.0+ Workflows (T2V, I2V/FL2V, R2V @ 480P & 768P) β€” β€” β€”
showcases/*.mp4 15-Second (362f @ 24fps + 32kHz Stereo Dialogue) Multi-Precision Comparisons β€” β€” β€”

πŸš€ Key Highlights

  1. True 3–4 Step Generation up to 15 Seconds (362 Frames @ 24fps + 32kHz Stereo Audio)
    • Eliminates single-frame I2V freeze, pure T2V collapse, and the severe mid-clip (5s–10s) double-exposure ghosting/blur seen in No-LoRA 4-step runs (6.5×–10Γ— higher Laplacian optical sharpness).
  2. One LoRA for All Base Precisions (Pure BF16 + ConvRot INT8 + GGUF Q4_K_M)
    • Distilled on 8Γ—NVIDIA B200 clusters with online student weight quantization perturbation (45% GGUF Q4_K_M, 35% Block-256 Hadamard ConvRot INT8, 20% Pure BF16) + Projected Distribution Matching Distillation (HyperPDMD) + 5-frame 3D-Conv VideoVAE spectral discriminator.
    • Runs under < 9 GiB VRAM on consumer GPUs with UnetLoaderGGUFAdvanced (patch_dtype="target") (0.00 GiB extra VRAM overhead!).
  3. 100% Native ComfyUI Compatibility (208/208 Layers)
    • Targets all 208 MiniMax-H3 DiT + TokenRefiner linear projections (comfy.lora.load_lora = 208/208 = 100.0%).

πŸ› οΈ Recommended ComfyUI 0.33.0+ Settings

  • Base Model:
    • < 9 GB VRAM: minimax-h3-{fl2va|ref2va}-pruned-Q4_K_M.gguf (UnetLoaderGGUFAdvanced, 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: 1.0 (0.95 ~ 1.05)
  • Sampler & Guider: SamplerCustomAdvanced + BasicGuider (CFG = 1.0) + KSamplerSelect ("euler")
  • Scheduler: BetaSamplingScheduler (steps = 4 or 3, alpha = 0.6, beta = 0.6)
  • MiniMaxH3SigmaShift:
    • FL2VA / I2VA / T2VA: shift_video = 12.0, shift_audio = 3.0
    • Ref2VA (1344Γ—768 Multi-Ref): shift_video = 9.0 (9.0 ~ 12.0), shift_audio = 3.0, and set ref_image_size = "max" on MiniMaxH3ReferenceToVideo (include a high-res chest-up portrait close-up on character turnaround sheets for pore-level 768P facial fidelity).
Downloads last month
1,405
Inference Providers NEW

Model tree for S3ckn11g0r/CROWN-MiniMax-H3-HyperPDMD-HD-3-4Step-LoRA

Adapter
(55)
this model