LSDJesus/MiniMax-H3-Image-fp8-nvfp4

MiniMax-H3-Image fp8 scaled and NVFP4 (ComfyUI)

Two more precisions of the MiniMax-H3-Image diffusion transformer for ComfyUI, alongside the bf16, int8 ConvRot and w4a8 files in Comfy-Org/MiniMax-H3-Image. They are drop-in replacements for those: same layout, same loader, same workflow. GGUF versions are in LSDJesus/MiniMax-H3-Image-GGUF.

Modified files. These are quantized versions of MiniMax-H3-Image, made from Comfy-Org's minimax_h3_image_pruned_bf16.safetensors. MiniMax-H3-Image is licensed under the MiniMax-H3-Image Community License Agreement, Copyright © 2026 MiniMax. All Rights Reserved. See LICENSE and NOTICE. Among other terms, the license grants rights only outside the European Union, the United Kingdom, South Korea and the United States (sections I.5 and II), and it includes an acceptable use policy. Read it before downloading.

Files

File	Format	Size	Runs natively on	Speed
minimax_h3_image_pruned_fp8_scaled.safetensors	fp8 e4m3, one scale per weight tensor	21.0 GB	RTX 40 and RTX 50	2.1 steps/s
minimax_h3_image_pruned_nvfp4.safetensors	NVFP4 (fp4 with fp8 scales per 16 values)	12.5 GB	RTX 50 and other Blackwell	3.4 steps/s

Speed is sampling at 2048×2048 on an RTX PRO 6000 Blackwell. On the same card Comfy-Org's int8 ConvRot ran 2.6 steps/s and w4a8 2.5. On cards without native support for a format, ComfyUI still runs it, more slowly.

Requirements

• ComfyUI with MiniMax-H3-Image support, merged on 10 November 2026 (ComfyUI 0.44).
• The text encoder (qwen3vl_32b_minimax_h3_nvfp4_awq or _int8_convrot) and the VAE (minimax_h3_image_vae_fp16) from Comfy-Org/MiniMax-H3-Image.

Put the file in models/diffusion_models/, open ComfyUI's MiniMax H3 Image template and choose it in Load Diffusion Model.

What is quantized

The same 200 layers that Comfy-Org quantizes in its int8 and w4a8 files: the fused qkv_proj, out_proj, mlp.fc1 and mlp.fc2 of the 50 transformer blocks, 95.8% of the weights. The token refiner, the condition projection, the adaLN time-curve basis, the patch and output projections and all norms are kept as in the bf16 source, as are the per-block int8 attention setting and the config metadata with the 6-step PDD schedule.

• fp8 scaled: e4m3 weights with one scale per tensor. mlp.fc2 is tagged full_precision_matrix_mult (fp8 weights, matrix multiply in bf16), as in Comfy-Org's own MiniMax-H3 fp8 files. Its input has outliers that ComfyUI's on-the-fly per-tensor fp8 activation scale cannot hold: without the tag the render drifted far more than NVFP4 (the subject changed and text was misspelled). The other layers quantize their activations on the fly; there are no calibrated activation scales.
• NVFP4: E2M1 values with an e4m3 scale per block of 16 and one global scale per tensor, made with ComfyUI's comfy-kitchen quantizer.

Quality

One render per file, seed 42, 2048×2048, 6 steps, compared with Comfy-Org's int8 ConvRot render. Left: whole image; right: a 1:1 crop. A 6-step model changes its composition a little with any change to the weights, so differences in pose and layout are normal; look for broken texture and text.

The template's own prompt:

A held-out photographic prompt (a fruit market with signs):