Midjourney ran its first models entirely inside a chat window

The process wall

Midjourney's image generator did not ship as a standalone application. It moved into open beta in July 2022 inside a Discord server, and stayed there for years: a maker typed a prompt into a chat channel and watched a four-image grid arrive from a bot. The company's own Legacy Features documentation, read as retrieved on 16 September 2026, records that Version 2 was the default model from April 2022 to July 2022 and that Version 3 was the default model from July 2022 to November 2022, placing the open beta at that V2-to-V3 handover. A homepage captured in November 2022 still describes the company as a small self-funded team of eleven full-time staff running the service through a public Discord.

What the documents show

The chat interface was not incidental. Every prompt, upscale and variation request was posted in a shared channel, making other members' prompts and outputs visible by default, a form of open, socially legible iteration a single-user desktop tool does not have. The current Version documentation still organises the tool's entire history by model version number, from V1 through V7 plus a separate Niji line, which shows Midjourney treats a version change as a distinct, dated capability shift rather than a silent update. Neither document names a training dataset or licence for the images the models learned from; the Legacy Features page describes only what each version could render.

Authorship and limitation

A prompt in a chat window is a description, not a specification, and the two are not the same: the model interprets tone, medium and composition words probabilistically, so identical prompts produce different grids each time. The documentation's own version-comparison chart, built to help a user pick a model, is itself an admission that outputs are version-dependent in ways a user must actively manage rather than assume fixed. Ownership and commercial terms for chat-generated work sit in Midjourney's separate terms of service, not in the version documentation, and are a contractual grant rather than a copyright ruling on the images; a maker relying on generated work commercially should check that document rather than infer rights from a feature description.

What survives

What is checkable from these pages is narrow: which version produced a given image, what that version's documented strengths and limits were at the time, and whether the work happened in a public or private channel. Editorially, a credit line for Midjourney output should record the version number, since the Legacy Features page treats successive versions as materially different tools, not cosmetic updates.

• Check the Legacy Features page for which version was active on the date an image was made.
• Read the current Version documentation for what its comparison chart does and does not claim.
• Ask whether a generated image was made in a public Discord channel or a private one, since prompt visibility differs.

The open beta's real innovation was social: a subscription wrapped around a chat log, where iteration was visible by default and version history became the closest thing to a changelog a maker could consult.