🚀 Midjourney Office Hours Highlights: February 5th

Midjourney's V7 model is making waves with its promising advancements. A month into its training, V7 is already showcasing significant improvements in image quality over its predecessor, V6. The team is currently running two training experiments: one with enhanced text data and another with less, aiming to finalize their choice by the end of the week. The focus remains on refining the model architecture, as the core dataset has remained largely unchanged since V4.

For a quick overview, head straight to the TL;DR; section otherwise keep on reading.

Extended Recap

⚡ Speed Optimization: A Key Focus

Speed is of the essence in AI image generation, and Midjourney is keenly aware of this. The final stage of V7's development involves a distillation process designed to optimize and accelerate the model. Faster iterations during image refinement are crucial for enhancing user experience. The team is meticulously fine-tuning compute settings and ensuring server readiness, with the distillation process expected to wrap up in a few days.

📊 Dataset Challenges and Future Plans

While a new, larger dataset was collected for V7, it wasn't cleaned in time for the release. Consequently, the current model relies on an older dataset, with plans to integrate the expanded, cleaned data in a future update. This enhancement is anticipated to better capture exotic details, such as mythological or historical elements, even if common elements remain largely unchanged. Future efforts will also focus on improved A/B testing to assess the impact of dataset updates.

🔮 Looking Ahead to V8

V7 was intentionally simplified compared to V6 to boost performance, but V8 might reintroduce complexity for further enhancements. The team is still deciding whether V8 will prioritize raw capacity or speed improvements. Architectural innovations are on the horizon, promising a radical leap in model performance.

🎥 Video Model Development: A New Frontier

Midjourney is also venturing into video model development, with promising proof-of-concept results shifting expectations. Video quality is anticipated to improve significantly within the year. The team is evaluating three potential approaches: building in-house video models, partnering with other companies, or a hybrid strategy. Key considerations include achieving consistency, managing costs, and addressing different server requirements. Integrating video into the overall workflow while balancing resource constraints is a priority.

🖼️ 3D Capabilities and Reframing Features

Exciting developments in 3D capabilities are underway, with successful experiments indicating that 3D generalization from Midjourney images is now achievable. Short-term goals include features for moving the camera, reframing, and re-rendering images. The long-term vision involves creating an open-world, real-time re-rendering experience. To accelerate development, the team plans to hire additional talent in the 3D area. Internal discussions are ongoing to balance refining core image generation with expanding into new modalities.

🌐 Website and User Interface Enhancements

Midjourney is also focusing on enhancing its website and user interface. The infrastructure has been updated to support larger batch processing and more advanced editing workflows. New interface concepts are being explored, including enhanced web editors with layers, improved selection tools, folders, and workspaces. Existing features like personalization, mood boards, and Sref codes will continue to be supported. However, there is some uncertainty about the exact timing for interface updates due to potential changes in V7’s batch size and performance.

TL;DR

🚀 V7 Model: Significant image quality improvements; final training choice soon. ⚡ Speed Optimization: Distillation process to enhance speed; crucial for user experience.

📊 Dataset: New data not ready for V7; future updates to include cleaned, expanded data.

🔮 V8 Plans: Simplified V7; V8 may reintroduce complexity for enhanced performance.

🎥 Video Development: Promising results; exploring in-house, partnerships, or hybrid strategies.

🖼️ 3D Features: Achievable 3D generalization; plans for camera movement and re-rendering.

🌐 UI Enhancements: Infrastructure updates; exploring advanced editing and personalization features.

Stay tuned for more updates as Midjourney continues to innovate and enhance its platform!

If you want to support me, feel free to buy me a coffee ☕️ 

If you’re not subscribed yet to my newsletter Imagine Weekly, I’d be thrilled to welcome you on board!