InvokeAI 7.0.0-alpha.3 Adds Role-Based Boards and New Models

What happened

InvokeAI 7.0.0-alpha.3 is an alpha release that introduces new features and improvements. The release includes board sharing by role, allowing users to be assigned Viewer, Contributor, Editor, or Owner roles. It also supports the Qwen-Image-2.1 model, including the Base model and the 8-step Turbo version.

The project file format has been changed to allow each member to have their own settings. This change means projects saved in alpha 3 cannot be opened in alpha 2, so users must stay on alpha 3. Additionally, a new opt-in setting for the SageAttention attention backend is included.

Why it matters

The release of InvokeAI 7.0.0-alpha.3 marks a shift in how users collaborate on projects by introducing role-based board sharing. This feature enhances team workflows by assigning specific permissions, which is particularly useful for larger teams or organizations. The inclusion of the Qwen-Image-2.1 model expands the tool's capabilities for image generation, offering users more options for creative tasks.

Key details

The SageAttention attention backend, designed for NVIDIA GPUs, improves generation speed by 15–35% on an RTX 4090 under Windows for 2048² resolutions. It works with multiple models, including FLUX.1, FLUX.2, and SDXL. Users are advised to create a new root directory for v7 to avoid conflicts with v6 databases, as upgrading an existing v6 database to v7 makes it incompatible with v6.

The release is described as alpha software with potential instability and incomplete features. Users are encouraged to report bugs via GitHub issues or Discord channels. The new version also changes the project file format, requiring users to stay on alpha 3 to maintain compatibility.