●Change8

v6.10.0

📦 invokeaiView on GitHub →
✨ 12 features🐛 2 fixes🔧 9 symbols

Summary

InvokeAI v6.10.0 introduces major support for the Z-Image Turbo model family and several quality-of-life improvements, including workflow tagging, prompt weighting hotkeys, and a new hotkey editor. This release also adds VRAM management options and PBR map generation capabilities.

Migration Steps

  1. To use Z-Image Turbo models, go to the Models tab and launch the "Z-Image Turbo" bundle from the Launchpad, or manually install individual models.
  2. When using quantized ZiT models (e.g., Q4_K), ensure the standalone Qwen3 encoder and a Flux VAE model are also installed (this is handled automatically when installing a ZiT starter model).
  3. To enable automatic VRAM release, add the configuration option `model_cache_keep_alive_min` to your `invokeai.yaml` file (set to 0 or omit to disable/use default).

✨ New Features

  • Introduced backend support for the Z-Image Turbo (ZiT) image generation model family, supporting various formats including diffusers, safetensors, GGUF, and quantized versions.
  • Added support for ZiT LoRA models, controlnet models, canvas functions, and regional guidance.
  • Implemented Workflow Tags feature allowing users to add arbitrary text tags to workflows for filtering in the Workflow Editor.
  • Introduced a new Prompt Template node in the Workflow Editor to apply built-in or custom prompt style templates to prompts.
  • Added hotkeys for adjusting prompt weighting: Ctrl-Up Arrow to increase weight and Ctrl-Down Arrow to decrease weight.
  • Introduced a new Hotkey Editor UI for assigning keyboard shortcuts to major UI functions.
  • Enabled bulk operations (currently Delete) in the Model Manager tab.
  • Added functionality in the Canvas to extract an arbitrary portion of all visible raster layers covered by the Inpaint Mask, compositing it as a new raster layer.
  • Added support for generating PBR maps (Normal, Displacement, and Roughness) from images via a context menu option.
  • Added new Heun and LCM schedulers for FLUX models (dev and schnell).
  • Added SDXL Color Compensation option to mitigate color drift during sequential SDXL image2image operations.
  • Added an option to automatically release GPU VRAM after a set period of inactivity via the `model_cache_keep_alive_min` configuration setting in `invokeai.yaml`.

🐛 Bug Fixes

  • Fixed issues where FP8 scaled weights models for ZiT produced image artifacts (though full support is still pending).
  • Fixed limitations where prompt weighting hotkeys did not work properly with numeric weights or prompts containing .add() or .blend() functions (to be fully fixed in the next point release).

🔧 Affected Symbols

Z-Image Turbo Model FamilyPrompt Weighting Hotkeys (Ctrl-Up/Down Arrow)Hotkey EditorModel ManagerCanvas Masked Area ExtractionPBR Map Generation FilterFLUX Model Schedulers (Heun, LCM)SDXL Color Compensation Optionmodel_cache_keep_alive_min (configuration option)