P-Image-Edit is a high-performance image-to-image editing model developed by Pruna AI, designed for real-time, instruction-based visual transformations. It allows users to perform complex editing tasks—such as object replacement, style transfer, and background modifications—using natural language prompts. The model is part of Pruna's "P-Model" series, which focuses on delivering high-fidelity generative outputs with extreme computational efficiency, typically achieving inference speeds of approximately one second.
The model is built upon an optimized Qwen-based architecture and refined using Pruna AI’s proprietary "smashing" techniques, which incorporate pruning, quantization, and distillation. These optimizations enable the model to maintain high prompt adherence and visual quality while significantly reducing the latency and cost compared to traditional large-scale multimodal models. P-Image-Edit is specifically noted for its ability to render accurate, styled text within images and its reliable spatial consistency during edits.
A distinguishing feature of P-Image-Edit is its support for multi-image referencing, allowing it to process up to five input images simultaneously. This capability enables advanced workflows such as merging elements from different scenes, applying specific styles from reference images, or performing complex multi-subject compositions. Official prompting guidelines suggest using descriptive adjectives and referring to specific inputs numerically (e.g., "image 1") to maintain precise control over the output.