P-H3-Image-Comfy: 8MP image generation with 21GB VRAM in 8 steps

A file set optimized to efficiently run the MiniMax-H3-based image generation and editing model in a ComfyUI environment. The original model's 143GB size is compressed to a 20GB download and a 21GB VRAM peak. It renders 8-megapixel resolution images in 8 steps without CFG.

1 / 5

The core feature is editing, which accepts up to 8 input images in addition to simple text generation. It includes a built-in prompt upsampler based on Qwen3.5-9B, automatically incorporating the process of converting short request sentences into detailed prompts into the workflow. Since it uses only ComfyUI's native nodes, it can be run immediately without installing custom nodes.

Files are provided in three precision levels: bf16, int8, and w4a8, which can be freely combined by component. The int8 version maintains almost the same quality as bf16, while the w4a8 version further reduces size, shortening the 2048x2048 image generation time from 19 seconds to 6.3 seconds in a 48GB L40S environment. If you already have MiniMax-H3, you can also reuse the existing transformer by applying only the 2GB-sized LoRA file.

It delivers optimal performance on GPUs with NVIDIA Ampere or higher, or AMD RDNA3/4 architecture. However, usage in the European Union, UK, South Korea, and US regions is restricted without a separate license under the terms, and the generated output must include the 'Powered by MiniMax H3' label. It is sensitive to prompts, so it is recommended to keep the upsampler in the workflow.