Change8

v5.12.0

Breaking Changes
📦 invokeaiView on GitHub →
1 breaking4 features🐛 6 fixes1 deprecations🔧 5 symbols

Summary

This release introduces support for Nvidia 50xx GPUs via a PyTorch bump, adds model relationship management, and implements stricter validation for node default values, which may require updates for custom node authors.

⚠️ Breaking Changes

  • Node default values are now strictly validated at application startup. If a node defines a field (e.g., ImageField) with a default value that does not match the field type (e.g., default=None), the application will error on startup. Fix by either updating the type annotation to include None (e.g., ImageField | None) if the field is optional, or by removing the invalid default or providing a valid default value that satisfies the field constraints (e.g., for an integer field with gt=10, the default must be > 10).

Migration Steps

  1. Node authors must review their custom nodes to ensure all InputField default values are valid according to their type annotations, or update type annotations to allow None if the field is optional. Nodes failing validation will cause application startup failure.
  2. Users are recommended to use the new Invoke Launcher for installation and updates, following the Quick Start guide.

✨ New Features

  • Support for Nvidia 50xx GPUs enabled by bumping PyTorch dependency to v2.7.0.
  • New model relationship feature allowing users to link models (primarily LoRAs to main models) in the model manager.
  • New IP Adapter methods: `Style (Strong)` and `Style (Precise)`.
  • Adds Imagen4 scaffold support.

🐛 Bug Fixes

  • Fixed GGUF quantization on MPS.
  • Addressed pydantic deprecation warning related to accessing `BaseModel.model_fields`.
  • Re-enabled new model classification API with safeguards.
  • Fixed transformers bug with SAM.
  • Improved GGUF performance with torch.compile.
  • Fixed node validation issues regarding pydantic field type massaging and better defaults parsing/error handling.

🔧 Affected Symbols

ImageFieldBaseInvocationInputFieldGGMLTensorSAM

⚡ Deprecations

  • The previous IP Adapter style method is renamed to `Style (Simple)`.