SearchingMan/P-H3-Image-GGUF

P-H3-Image GGUF

GGUF files of PrunaAI/P-H3-Image for stable-diffusion.cpp. The files are modified: they are converted from the published checkpoint, with tensors renamed and re-laid-out into MiniMax-H3's original layout, and quantized.

File	Component	Size	Types
p_h3_image-Q4_K_M.gguf	transformer (33B)	18.8 GB	attention and MLP Q4_K, block modulation Q4_0, the rest BF16
p_h3_image_text_encoder-Q8_0.gguf	Qwen3-VL-4B text encoder + text connector	5.4 GB	language model Q8_0, vision BF16, connector F32
p_h3_image_vae-F16.gguf	VAE with the image decoder	5.2 GB	F16

Usage

P-H3-Image support in stable-diffusion.cpp is pending upstream; until it is merged, build the p-h3-image branch. Text-to-image:

sd-cli --diffusion-model p_h3_image-Q4_K_M.gguf --llm p_h3_image_text_encoder-Q8_0.gguf --vae p_h3_image_vae-F16.gguf -p "A German Expressionist oil painting of a Berlin street at night, with gaunt figures hurrying past a glowing gas lamp." -W 2048 -H 2048 --diffusion-fa

Editing, with up to eight images passed as -r (the first is the image to edit):

sd-cli --diffusion-model p_h3_image-Q4_K_M.gguf --llm p_h3_image_text_encoder-Q8_0.gguf --vae p_h3_image_vae-F16.gguf -r street.png -p "Make it a snowy winter night: heavy snow falls across the whole scene and covers the pavement and rooftops. Keep the painting style, composition and figures unchanged." -W 2048 -H 2048 --diffusion-fa

• Sampling uses the checkpoint's 8-step sigma grid and no CFG.
• The model works best at 4 to 8 megapixels and with detailed prompts; the reference pipeline's prompt upsampler is not included.
• A 2048x2048 image takes about 45 s of sampling on an RTX 5090 Laptop GPU.

License

MiniMax H3 is licensed under the MiniMax H3 Community License Agreement, Copyright © 2026 MiniMax. All Rights Reserved. See LICENSE and NOTICE. The license excludes the European Union, the United Kingdom, the Republic of Korea and the United States: it grants no rights to use, reproduce or distribute these files there. The text encoder's Qwen3-VL-4B parts are under the Apache License 2.0 (LICENSE-APACHE).