Comfy-Org Brings MiniMax H3-Image's 4K Local Generation to ComfyUI
Comfy-Org has repackaged MiniMax's open-weight H3-Image model into quantized single-file checkpoints that drop straight into ComfyUI workflows.
- Comfy-Org published a repackaged, quantized MiniMax H3-Image build ready to drop into ComfyUI
- Repack ships int8 DiT, NVFP4 AWQ Qwen3-VL-32B text encoder, and fp16 VAE as single files
- Underlying MiniMax H3-Image is a 33.1B open-weight model generating 4096x4096 images
- Six sampling steps with no classifier-free guidance, up to 16 reference images supported
- Shares text encoder with the existing Comfy-Org MiniMax H3 video repack
- Community license excludes users in the US, EU, UK, and South Korea
Comfy-Org has repackaged MiniMax H3-Image as three quantized safetensors files arranged for ComfyUI’s standard model folders. The release preserves the upstream model while reducing its storage and memory demands, removing much of the manual conversion work required by the original multi-file checkpoint.
MiniMax H3-Image is a 33.1 billion-parameter, open-weight diffusion transformer derived from MiniMax’s H3 video architecture. It supports text-to-image generation, reference-conditioned generation, and image editing. MiniMax reports native output around 16 megapixels, including 4096 by 4096 pixels, with up to 16 reference images and a six-step sampling schedule that does not use classifier-free guidance.
The package at a glance
| Component | Packaged format | Purpose |
|---|---|---|
| Diffusion transformer | Pruned int8 with ConvRot | Denoises the latent representation to generate the image |
| Qwen3-VL-32B encoder | NVFP4 with AWQ | Encodes prompts and reference-image instructions |
| Image VAE | FP16 | Converts between image pixels and the model’s latent representation |
The ComfyUI package contains one safetensors file for each component:
minimax_h3_image_pruned_int8_convrot.safetensorsqwen3vl_32b_minimax_h3_nvfp4_awq.safetensorsminimax_h3_image_vae_fp16.safetensors
Quantization lowers memory use by storing weights at reduced precision. The main transformer uses int8 weights, while the Qwen3-VL encoder uses NVIDIA’s 4-bit NVFP4 format with activation-aware weight quantization. Comfy-Org has not published an exact VRAM requirement, so practical limits will depend on resolution, attention settings, offloading, GPU architecture, and support for the required quantization kernels.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
