DALL·E 2

Generates photorealistic images from natural language descriptions using a diffusion-based generative model trained on large-scale image-text pairs. The system uses a two-stage architecture: first, a CLIP-based text encoder converts natural language prompts into a learned embedding space; second, a diffusion decoder iteratively denoises random noise conditioned on these embeddings to produce high-fidelity 1024×1024 pixel images. The model employs classifier-free guidance to balance prompt adherence with image quality.

Enables selective editing of images by masking regions and regenerating only the masked areas while preserving surrounding context. The system uses a masked diffusion process where the model conditions on both the original unmasked pixels and the text prompt, iteratively denoising only the masked region. Outpainting extends this to generate new content beyond image boundaries, effectively expanding the canvas while maintaining visual coherence with existing content.

Generates multiple diverse variations of a provided image while maintaining core visual characteristics (composition, style, subject matter). The system encodes the input image into the CLIP embedding space, then uses the diffusion model to generate new images conditioned on this embedding with added noise, producing semantically similar but visually distinct outputs. This enables exploration of design alternatives without requiring new prompts or manual iteration.

Provides REST API endpoints for programmatic image generation, enabling integration into applications, workflows, and batch processing pipelines. Requests are submitted asynchronously with prompt, size, and quantity parameters; responses include image URLs and metadata. The API supports rate limiting, quota management, and usage tracking, allowing developers to build scalable image-generation features without managing model infrastructure.

Implements automated content filtering and policy enforcement to prevent generation of prohibited content (violence, sexual material, copyrighted works, etc.). The system uses a combination of text-based prompt filtering (detecting policy violations in input prompts) and image-based filtering (detecting policy violations in generated outputs) before returning results to users. Violations are logged and may result in account restrictions.

Supports generation of images at multiple resolutions (256×256, 512×512, 1024×1024 pixels) to accommodate different use cases and cost constraints. The underlying diffusion model is trained to handle variable resolutions through resolution-aware conditioning, allowing users to trade off image quality and detail against generation time and API costs. Smaller sizes generate faster and cost less; larger sizes provide higher fidelity.

Returns the 'revised prompt' used for generation alongside generated images, showing how the system interpreted or modified the user's input prompt. This transparency mechanism helps users understand how their natural language descriptions were processed, disambiguated, or adjusted by the model before image generation. Revised prompts are particularly useful when the original prompt was ambiguous or when the model made assumptions about the user's intent.