Midjourney

last edited

Midjourney is an image and video generator from a small independent lab in San Francisco. Founder David Holz previously co-founded Leap Motion. The open beta of Midjourney began in July 2022. Holz said in August 2022 that the lab was already profitable. Its models are closed, unlike many open diffusion models.

Prompting and references

Users send a prompt on the website or through a Discord bot. Each job returns a grid of four images. Any of them can be varied, edited, or passed on for upscaling. The web editor arrived in August 2024, alongside V6.1.

Reference images steer a result. A style reference (--sref) maps palette, texture, and mood from an uploaded picture onto new images. Character and omni references keep a person or object consistent across images. Personalization builds a taste profile from about 200 image ratings and has been on by default since V7. A separate Niji line is tuned for anime.

Early models

The first model appeared in February 2022, months before the public beta. From V4 onward, training ran on Google TPUs. V5 arrived as an alpha in March 2023. V6 followed in December 2023. Each version was tested as an alpha before it became the default.

V7 and V8

The V7 alpha opened in April 2025 and added Draft Mode for cheap, rough drafts. The V8 alpha came in March 2026 on rewritten code. The lab moved from TPUs to a GPU-native PyTorch stack, giving about five times faster jobs and native 2K output.

In July, the V8.2 release became the default, with a focus on aesthetics and personalization. An edit model followed in August. It takes text instructions and up to four reference images, and it handles inpainting and outpainting.

Video

Midjourney released its first video model, V1, in June 2025. It animates a still image rather than starting from text alone. Clips begin at 5 seconds and can be extended to a maximum of 21. Output is 480p and can loop.

Sources