OpenAI releases DALL·E 2

OpenAI released DALL·E 2, a text-to-image model offering higher resolution and greater photorealism than the original DALL·E, alongside inpainting (editing specific regions of an existing image) and the ability to generate variations of an uploaded image. Rather than a public launch, OpenAI opened access to roughly 400 trusted users first, with a waitlist for wider access that it expanded over the following months.

The accompanying system card documented a set of restrictions unusual for a consumer product at the time. Input filters blocked prompts involving sexual content involving minors, violence, and explicitly political or toxic content, though OpenAI acknowledged the filters had “known deficiencies” and could not catch every context-dependent misuse. Uploaded images of real people required the uploader to hold rights or consent, commercial use was disallowed during the preview, users had to disclose AI-generated content when sharing it, and every output image carried a visible watermark. Generation was rate-limited and subject to human review for policy violations.

DALL·E 2’s cautious rollout — small trusted cohort, waitlist, heavy filtering, mandatory watermarking — reflected OpenAI’s approach to releasing a capability it judged could be misused for disinformation or non-consensual imagery. That approach set a contrast that became sharper a few months later, when Midjourney opened a public beta on Discord with far fewer restrictions and Stability AI released Stable Diffusion’s weights outright, and the gap between the two strategies became one of the field’s recurring disputes.