More than six months after the last update, Midjourney has released a new image model - but it's a smaller jump.
Midjourney CEO David Holz announced version 6.1 of the company's image model via Discord. This new version builds on V6, released in late December, and promises more coherent, detailed, and visually appealing results.
According to the changelog, V6.1 delivers significantly better image quality overall, with fewer artifacts and improved textures. Holz highlights improvements in skin and 8-bit pixel art. Standard image generation is 25% faster.
The update focuses on refining well-known problem areas such as human extremities (arms, legs, and hands), as well as plant and animal representations. Small image details such as eyes, small faces, and distant hands are now more detailed and accurate. The update also introduces new upscaler tools.
Text accuracy has been improved, especially when drawing words with quotation marks in prompts. In a quick comparison, V6.1 tended to perform better, but was still not flawless.
A new parameter, --q 2, takes 25% longer to generate images but can add more texture at the expense of image coherence.
With V6.1, Midjourney is expanding the personalization features it introduced in June. The new model has improved nuance, surprise, and accuracy, though Holz doesn't explain exactly how this manifests itself in use. Users can now use personalization codes from older jobs, thanks to versioning.
Holz illustrates the challenge of articulating improvements in an AI image model with his final point on the list of changes: "Things should look 'generally more beautiful' across the board."
Midjourney V6.1 does not currently offer an inpainting/outpainting model, so features like zoom, reframe, repaint, and vary region still need the V6.0 model.
To maximize data collection and support future development, V6.1 is now the default model. Older models can be enabled as usual via /settings in the Discord bot or the settings menu on the website.
While only about two months elapsed between V5 and V5.1, Midjourney took over seven months for this update. Does this suggest that the development curve for improving diffusion models has flattened? It certainly seems like it's slowed down quite a bit after we saw insane progress in a year or so from the first Midjourney releases to v4.
According to a March announcement, V7 was slated for a summer release. But apparently the new features weren't enough to make the jump. V6.2 is expected to be released in August, with improvements based on data collected from users' experiences with V6.1.
