ComfyPilotComfyPilot
模型库 / 整模型 / MiniMax-H3-Image Exact-adaLN NVFP4
MiniMax-H3-Image Exact-adaLN NVFP4

MiniMax-H3-Image Exact-adaLN NVFP4

3057968

整模型 32 次下载 4 喜欢 更新于 2026-11-20

底模:MiniMax H3 Image

quantizedanimenvfp4

下载模型 在源站打开

文件清单 线路:原始(huggingface.co) · 切换

下载后放进对应的 ComfyUI 目录。models/ 指 ComfyUI 的模型根目录。

权重 minimax_h3_image_exact_adaln_nvfp4.safetensors 11.7 GB 下载

说明

MiniMax-H3-Image Exact-adaLN NVFP4

作者:lepetitprince716355

◆ MiniMax-H3-Image · Exact-adaLN NVFP4 (12.61 GB) · Zero-ULP Exact 7-Step Modulation + Native Blackwell FP4 A 12.61 GB (11.74 GiB) Blackwell NVFP4 checkpoint of MiniMax-H3-Image (33.1B DiT) for ComfyUI that replaces the lossy 8-dimensional F16 curve approximation found in standard pruned builds with a zero-error 7-step one-hot lookup table computed directly from the official 26.0 GB adaLN MLPs.

What Was Changed

  • 1. Lossless Exact 7-Step One-Hot adaLN Table (0 BF16 ULP Error): Upstream MiniMax-H3-Image spends 26.0 GB (39.5% of the 66.3 GB DiT) on per-block adaLN MLPs (2688 -> 96768 across 50 blocks). Standard pruned checkpoints compress those MLPs into a [1025, 8] F16 basis curve, introducing p99 = 5 BF16 ULP error across blocks (up to 7 ULP on step 0 and 9 ULP on visual reference rows). Because H3-Image's 6-step PDD solver only ever queries the DiT at 7 fixed timesteps (t = 0, 0.0324763, 0.1124515, 0.25, 0.4672261, 0.7679911 plus 0.999 for reference images), we pack a [1025, 7] nearest-neighbor one-hot table into ComfyUI's native curve format alongside exact FP64-accumulated official modulation vectors (~136 MB total). ComfyUI loads it out-of-the-box with zero code patches and 0 BF16 ULP error across all 50 blocks.
  • 2. Blackwell Native NVFP4 DiT Projections (comfy-kitchen TensorCoreNVFP4Layout): Quantizes the 200 heavy linear projections (blocks.0..49.{attn.qkv_proj, attn.out_proj, mlp.fc1, mlp.fc2}) to NVFP4 (E2M1 values with F8_E4M3 block scales per 16 elements + F32 per-tensor scale). Keeps token_refiner, condition_proj, video_patch_proj, final_layer.video_out (6-step PDD head bank), all q/k/norm weights, and the exact adaLN table in full BF16 / FP32.

Measured Footprint & Speed (4 MP · 6 Steps)

  • File Size: 12.61 GB (12,605,496,924 bytes, SHA-256: e5a9122b14dc31ab4da67b1a7513041afb3ced7d31dd81f501c1aa16b958aa2a) vs 40.30 GB Pruned BF16 (3.2x smaller) and 66.28 GB unpruned DiT (5.3x smaller).
  • Resident DiT VRAM: 11.75 GiB (12.34 GiB peak DiT VRAM at 4 MP text-to-image; 14.00 GiB peak on dual-reference 4 MP editing — fits 16 GB Blackwell GPUs such as RTX 5080 / 5070 Ti when paired with text-encoder offloading or qwen3vl_32b_minimax_h3_nvfp4_awq).
  • 6-Step 4 MP DiT Time: ~10.0 - 10.5 s (~1.65 s/step at 1664x2432 / 2048x2048, 1.42x faster than Exact-adaLN BF16).

Recommended ComfyUI Setup (Exact PDD ManualSigmas)

  • Requirements: ComfyUI >= 0.45.0, comfy-kitchen >= 0.2.41, PyTorch with cu130+. Native 4-bit Tensor Core speedup requires NVIDIA Blackwell GPUs (RTX 50-series / GB-series); older GPUs dequantize weights on the fly.
  • Text Encoder (CLIPLoader, type: minimax): qwen3vl_32b_minimax_h3_bf16.safetensors (pass --bf16-text-enc so ComfyUI does not downcast BF16 to FP16) or qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors.
  • VAE (VAELoader): minimax_h3_image_vae_fp32.safetensors (--fp32-vae) or minimax_h3_image_vae_fp16.safetensors.
  • Sampling: SamplerCustomAdvanced + KSamplerSelect (euler) + BasicGuider + ManualSigmas (6 PDD steps, no CFG). Drag any showcase PNG from this page directly into ComfyUI to load the exact workflow!

1.0, 0.9675236940383911, 0.8875485062599182, 0.75, 0.5327739119529724, 0.23200887441635132, 0.0Why ManualSigmas instead of KSampler (simple, 6 steps)? ComfyUI's simple 6-step scheduler lands at fractional grid indices on 4 of the 6 steps, drifting up to 8.7e-4 from the official PDD sigma schedule. Feeding the 7 values above into ManualSigmas matches the official PDD schedule to < 1.5e-8. License & Attribution

  • Derivative of MiniMaxAI/MiniMax-H3-Image and Comfy-Org/MiniMax-H3-Image, governed by the MiniMax-H3-Image Community License Agreement (note upstream territory exclusions in Section I.5: EU, UK, Republic of Korea, USA; full text in LICENSE-MODEL.md in the linked repositories). Not an official MiniMax release.

示例图