krea2: CREPA self-flow, LayerSync, TREAD, block swap, cfg-zero star, skip layer guidance, and flowmap generation (AnyFlow)

[bghira]

This pull request adds new experimental features and improves support for the Krea 2 model in both the backend logic and the documentation. The most significant changes are the addition of advanced training and inference options for Krea 2, including TREAD training, CREPA self-flow, signed-timestep and FlowMap conditioning, and new classifier-free guidance strategies. Documentation in all supported languages has been updated to reflect that Krea 2 is now compatible with Musubi block swapping.

Krea 2 Model Enhancements:

• Added support for TREAD training to Krea2, including a tread_init method that configures TREAD routes and router initialization.
• Implemented CREPA self-flow support in Krea2, including methods for batch preparation and hidden state selection.
• Enabled experimental timestep conditioning for Krea 2, including signed-timestep (TwinFlow) and FlowMap delta-timestep support in the transformer and model logic. [1] [2] [3]
• Added logic to support Musubi block swap defaults when loading pretrained weights for Krea 2. [1] [2]

Inference Pipeline Improvements:

• Added new classifier-free guidance strategies (cfg_zero_star), skip-layer guidance options, and an optimized scaling function to the Krea 2 pipeline, allowing for more advanced and configurable denoising behavior during inference. [1] [2] [3]

Documentation Updates:

• Updated the --musubi_blocks_to_swap option in all documentation languages to include Krea 2 as a supported model for Musubi block swapping. [1] [2] [3] [4] [5] [6]

These changes collectively improve Krea 2’s flexibility for research and advanced users, expand its compatibility with memory-saving features, and ensure documentation is accurate and up-to-date.