Original title: PrunaAI/P-H3-Image
Pruna AI's image model, fine-tuned from MiniMax-H3, uses a single diffusers pipeline for both text-to-image generation and editing with 1–8 input images: the 1st image is the one to be edited, while the others serve as references (referred to in the prompt as Picture 2, Picture 3, etc.). Officially, it can generate 8MP images in 8 steps without CFG, excelling in realistic photography, cinematic visuals, editing-oriented styles, and graphic design featuring short quoted text. By default, the pipeline first runs a prompt expansion model based on Qwen3.5-9B, rewriting a one-sentence request into a full description that the model handles best; this can also be disabled to render the original text as-is. The weights include a 33.1B-parameter transformer, a 4B text encoder, a VAE, and the expansion model, totaling approximately 100GB in bf16.
Why it matters No CFG or negative prompts required; built-in prompt expansion allows generating complete scenes from one-sentence requests; additionally, there is a low-VRAM ComfyUI version
Rising score 109 · Likes 284 · Downloads 1666
Reasons
- Multi-image composition: Provide up to 8 images and a one-sentence prompt; the expander clarifies what to take from each image before compositing them into one. When editing, it specifies "what to keep" and "what to change," avoiding full redraws.
- Official prompting guidelines: Start by stating what the image is (e.g., "a street photograph"), describe only one scene, quote any text that should appear verbatim, and avoid exaggerated terms like stunning or 8k.
- Speed: Officially claimed to generate a 8MP text-to-image output in approximately 9 seconds on a single B200 (excluding expansion time); prompt expansion requires an additional 6–11 seconds.
Usage
- After running pip install git+https://github.com/huggingface/diffusers, use ModularPipeline.from_pretrained("PrunaAI/P-H3-Image"); passing the image parameter enables editing mode.
- For text-to-image, 8MP dimensions are recommended (default 2912×2912); for editing, use 4MP or 8MP. If VRAM is insufficient, ComponentsManager automatically unloads components, with peak usage around 76GB on an 80GB card.
- For ComfyUI, use the int8 / 4 bit files from PrunaAI/P-H3-Image-Comfy; officially claimed to run with 21GB of VRAM.
Risks
- The license is the MiniMax H3 Community License: it does not cover the EU, UK, South Korea, or the US (separate permission from MiniMax is required), and products must be labeled "Powered by MiniMax H3."
- Full configuration at 8MP with expansion requires approximately 101GB of VRAM; at 2MP and 4MP, faces, hands, and small details are more prone to errors, and small text or large blocks of text may still be rendered incorrectly.
- This entry has not been run; refer to the model card examples for actual results.