Woovee/MiniMax-H3-Image-LowVRAM

MiniMax H3 Image — low VRAM files for Woovee Studio

Selected, unmodified Comfy-Org repackaged weights of the official MiniMax H3 Image, released November 10, 2026. This is the official six-step image model, distinct from Pruna P-H3-Image and from the H3 video models. Powered by MiniMax-H3-Image.

Files

File	Size	Purpose
diffusion_models/minimax_h3_image_pruned_w4a8.safetensors	12.55 GB	Pruned image transformer, W4A8
text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors	15.69 GB	Original 32B encoder, NVFP4 AWQ
vae/minimax_h3_image_vae_fp16.safetensors	1.17 GB	Image VAE

Total download: approximately 29.41 GB. Exact upstream revision and SHA256 values are in source-manifest.json. These quantized files are copied without modification; the selection and documentation are maintained by Woovee.

Usage

Use ComfyUI commit 989d93d0d968e704c589ce0146bb18c77f6b50cc or a compatible later release. Place each file in its matching ComfyUI/models/ folder. Use MiniMaxH3ImageConditioning, KSampler with 6 steps, CFG 1, euler, simple, and the included image VAE. Follow the official ComfyUI templates.

Start at 2048 × 2048 (4 MP) for speed. The upstream model recommends around 16 MP for best detail, at greater memory and compute cost. Describe the entire image in detailed English. For editing, name references <Picture 1>, <Picture 2>, etc., and specify both the desired change and what should remain. The Studio's optional prompt enhancement is separate; a large prompt expander is not included in this weight set.

The deployment targets 24–32 GB NVIDIA GPUs with dynamic offloading, loading the text encoder and diffusion model in stages. Runtime overhead, reference count, resolution, GPU kernels and CPU RAM also matter. W4A8/NVFP4 outputs can differ from full precision.

Measured on November 10, 2026 with six steps and 2048 × 2048 output:

Test	Workflow time	Total handler time	Sampled peak VRAM
RTX 5090 (32 GB), first text-to-image	46.90 s	78.26 s	28,162 MiB
RTX 5090 (32 GB), warm one-reference edit	23.44 s	24.25 s	30,727 MiB
RTX 4090 (24 GB), first one-reference edit	26.35 s	50.61 s	22,975 MiB

All returned valid images; edits changed a blue teapot to green or orange while retaining composition. Workflow time includes component loading and output encoding; total handler time includes model hash verification and server startup. Provider queue and container startup are excluded. These are individual observations, not throughput guarantees or minimum-memory requirements. 24 GB and 32 GB tests passed at 4 MP with one reference. Higher resolutions, four simultaneous references and the other configured GPU types have not been benchmarked.

License and access

Read the included MiniMax-H3-Image Community License Agreement, which differs from the older H3 license. It excludes the EU, UK, Republic of Korea and USA under its standard grant. Separate prior written authorization from MiniMax is required for third-party hosted services or annualized revenue above USD 1 million. This repository distributes weights; it does not grant a hosted-service license or override the upstream terms. Do not use the model outside the license's applicable territory without authorization. The Studio deployment is intended for its owner's private use.

The original model, architecture and license belong to MiniMax; Comfy-Org produced these repackaged quantized files. See LICENSE and NOTICE.