
Meta's AI research tool that generates images from text prompts and freeform sketches
Make-A-Scene is a text-to-image AI generation system developed by Meta AI Research that combines text prompts with optional freeform sketches for greater creative control. The model uses a scene-based approach with human priors, producing high-fidelity images up to 2048x2048 pixels while enabling scene editing, text editing with anchor scenes, and story illustration generation.
Generates high-fidelity images from natural language text prompts using a transformer-based architecture
Accepts optional hand-drawn sketches alongside text prompts to give creators more control over composition and layout
Uses scene layouts as an implicit control mechanism, allowing users to specify elements, arrangements, depth, and structure
Produces images up to 2048x2048 pixels, a four-fold increase over initial capabilities
Enables editing specific parts of a generated scene while preserving the rest of the composition
Generates coherent sequences of images suitable for illustrating stories or narratives
Employs specialized knowledge over key image regions like faces and salient objects for improved image quality
Artists can sketch rough layouts and let the AI fill in detailed imagery, maintaining precise control over scene composition
Create coherent image sequences for stories or comics by using scene anchoring to maintain visual consistency across frames
Researchers and AI enthusiasts can explore scene-based generation techniques and build upon Meta's open research
Designers can quickly generate visual concepts by combining rough sketches with text descriptions for faster ideation
Allows modifying the text prompt while keeping the scene layout anchored for consistent variations

OpenAI's AI image generator built into ChatGPT for effortless creation