◆ 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.
Hugging Face Mirror & Checksums: https://huggingface.co/kefallrain/MiniMax-H3-Image-Exact-adaLN-NVFP4
GitHub Implementation, Exact-adaLN Build Scripts & ComfyUI Workflows: https://github.com/lepetitprince716-prog/minimax-h3-image-exact-adaln-nvfp4
What Was Changed
1. Lossless Exact 7-Step One-Hot
adaLNTable (0BF16 ULP Error): UpstreamMiniMax-H3-Imagespends26.0 GB(39.5% of the 66.3 GB DiT) on per-blockadaLNMLPs (2688 -> 96768across 50 blocks). Standard pruned checkpoints compress those MLPs into a[1025, 8]F16basis curve, introducingp99 = 5BF16 ULP error across blocks (up to7ULP on step 0 and9ULP 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, 0.7679911plus0.999for reference images), we pack a[1025, 7]nearest-neighbor one-hot table into ComfyUI's native curve format alongside exactFP64-accumulated official modulation vectors (~136 MBtotal). ComfyUI loads it out-of-the-box with zero code patches and 0 BF16 ULP error across all 50 blocks.2. Blackwell Native
NVFP4DiT Projections (comfy-kitchenTensorCoreNVFP4Layout): Quantizes the 200 heavy linear projections (blocks.0..49.{attn.qkv_proj, attn.out_proj, mlp.fc1, mlp.fc2}) toNVFP4(E2M1values withF8_E4M3block scales per 16 elements +F32per-tensor scale). Keepstoken_refiner,condition_proj,video_patch_proj,final_layer.video_out(6-step PDD head bank), allq/k/normweights, and the exactadaLNtable in fullBF16/FP32.
Measured Footprint & Speed (4 MP · 6 Steps)
File Size:
12.61 GB(12,605,496,924bytes, SHA-256:e5a9122b14dc31ab4da67b1a7513041afb3ced7d31dd81f501c1aa16b958aa2a) vs40.30 GBPruned BF16 (3.2xsmaller) and66.28 GBunpruned DiT (5.3xsmaller).Resident DiT VRAM:
11.75 GiB(12.34 GiBpeak DiT VRAM at 4 MP text-to-image;14.00 GiBpeak on dual-reference 4 MP editing — fits 16 GB Blackwell GPUs such as RTX 5080 / 5070 Ti when paired with text-encoder offloading orqwen3vl_32b_minimax_h3_nvfp4_awq).6-Step 4 MP DiT Time:
~10.0 - 10.5 s(~1.65 s/stepat1664x2432/2048x2048,1.42xfaster than Exact-adaLN BF16).
Recommended ComfyUI Setup (Exact PDD ManualSigmas)
Requirements: ComfyUI
>= 0.45.0,comfy-kitchen >= 0.2.41, PyTorch withcu130+. 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-encso ComfyUI does not downcast BF16 to FP16) orqwen3vl_32b_minimax_h3_nvfp4_awq.safetensors.VAE (
VAELoader):minimax_h3_image_vae_fp32.safetensors(--fp32-vae) orminimax_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'ssimple6-step scheduler lands at fractional grid indices on 4 of the 6 steps, drifting up to8.7e-4from the official PDD sigma schedule. Feeding the 7 values above intoManualSigmasmatches 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.mdin the linked repositories). Not an official MiniMax release.

