Midjourney V8: Speed Boosted, Premium Costs Rising

Midjourney recently unveiled a preliminary version of its V8 model, which is currently in community testing. This new model stands out for its ability to generate images approximately five times faster than previous versions. Among the new features, the --hd mode allows for the production of images in native 2K resolution, meeting the expectations of users seeking high visual quality.

The V8 model is designed to follow detailed instructions with increased accuracy, producing more coherent images and improving text rendering within images. Midjourney claims that this model is significantly better at understanding individual aesthetics through customization, style references, and moodboards, which can be transferred from V7. However, despite these improvements, the model remains based on a 1000% diffusion architecture, making it less effective than autoregressive models for complex prompts.

Midjourney has made this version available on its Alpha site, encouraging users to test it and provide feedback. Although V8 is presented as a fundamentally new model, it may require novel prompting strategies to fully leverage its capabilities. Users can experiment with various parameters such as --chaos, --weird, --exp, and --raw, and the model supports multiple aspect ratios.

Despite these advancements, the V8 model still shows limitations when faced with complex prompts. In tests with abstract concepts, such as a horse riding an astronaut, V8 produced results inferior to autoregressive models like Google’s Nano Banana and OpenAI’s GPT image 1.5. Even Midjourney's direct competitor, Flux, achieved better results in some cases.

Regarding costs, the premium features of V8, such as --hd, --q 4, or the use of moodboards, are currently four times slower and more expensive than standard work. The relax mode, which allowed for generating images at a lower cost, is not available at this time. Midjourney is working on a new server cluster to reintroduce this option and on more economical rendering modes. These cost increases could influence user decisions, particularly for those seeking precise prompt control at a lower cost.

The diffusion model of Midjourney V8, while improved, still lags behind hybrid architectures that integrate autoregressive components. The latter, used by models like Google’s Nano Banana and OpenAI’s GPT image 1.5, offer better prompt accuracy. Early tests show that V8 has not yet bridged this gap, particularly when executing complex prompts like the abstract concept of a horse riding an astronaut. Even Flux, a direct competitor, has demonstrated slightly superior performance in certain scenarios.

In conclusion, while Midjourney V8 brings significant improvements in terms of speed and image quality, the limitations related to its diffusion architecture and the increased costs of premium features may restrict its adoption by users demanding precise and economical prompt control.