DALL-E 2 turned inpainting into a waitlist product

The process wall

In April 2022, OpenAI showed a system that generates images from a written description and, unusually, edits an existing image rather than only inventing a new one. The launch page describes three separate operations: generating an image from a caption, producing variations of an uploaded image, and making an edit that adds and removes elements while accounting for shadows, reflections and textures. For a maker used to compositing by hand, that third operation is the interesting one: it treats a picture as an editable object rather than a single finished output, which changes what a draft looks like in this kind of workflow.

What the documents show

The research paper names the underlying method: a prior model turns a text caption into a CLIP image embedding, and a diffusion decoder turns that embedding into pixels. This two-stage design is why editing and variation exist at all, since the system manipulates a shared embedding space rather than running a caption through a single black box. When OpenAI opened a paid beta on 20 July 2022, it added an Edit tool built on that mechanism and stated it had removed the most explicit content from the training data and added filters against violent, adult and political prompts. Neither document names the training dataset itself; the beta announcement adds that users now get full usage rights to commercialize the images they create, including images made during the earlier research preview.

Authorship and limitation

The generative step is automated, but selection is not: a maker chooses a prompt, evaluates a grid of outputs, decides whether an edit reads as a correction or a fabrication, and decides what to publish. The beta page frames Edit as producing realistic and context-aware edits, a claim about plausibility, not about accuracy to any real scene. Content filters and the stated limits on generating recognisable public figures are OpenAI's own policy choices, not properties a viewer can verify from an image alone. The commercial rights grant in the beta announcement is a licence term OpenAI set, not a copyright determination, and it governs the account holder's use of outputs rather than any underlying claim about the images' status.

What survives

What a maker keeps from this record is narrow but usable: the announcement's own description of what each mode does, the credit and pricing structure it introduces, and the safety limits stated at the time. Editorially, a caption-to-image credit line should note which mode, generate, edit or variation, produced an image, since the mechanism differs between them.

• Read the launch page's own account of the caption-matching and photorealism comparisons it reports.
• Check the beta announcement for the exact commercial usage rights granted to an account.
• Read the paper's abstract for the prior-and-decoder architecture before describing the tool as one model.

Neither document supports a stronger claim than this: a text-and-image embedding was made editable, and OpenAI's own policies, not an independent audit, define what that editing can currently touch.