Google DeepMind explains in a new post how to use the improved image generation in Gemini to its full potential. Product Manager Naina Raisinghani shared specific prompting strategies to achieve better results with the updated model.
The company recommends including six key elements in prompts: subject, composition, action, location, style, and editing instructions. Users should be specific when describing subjects and direct when requesting modifications.
Google outlined five main techniques for effective use. Users can maintain character consistency across multiple images by establishing clear details in initial prompts. The system allows precise edits through conversational commands without regenerating entire scenes.
Creative composition enables blending multiple concepts into single images. Style transfer capabilities can completely change an image’s aesthetic while preserving the original subject. The model can also use logical reasoning to predict realistic outcomes or build complex scenes from simple concepts.
Google acknowledged current limitations including inconsistent stylization, text rendering issues, and aspect ratio maintenance problems.
