Experimental Directory
This directory contains experimental code, examples, and approaches that aren't currently used in the main nodes but might be useful for future development.
Contents
injection_methods.py
Alternative approaches to activation injection that work at different levels:
- LatentSpaceInjector: Manipulates latent space between denoising steps
- CrossAttentionHijacker: Intercepts cross-attention layers
- GuidanceModulator: Modifies classifier-free guidance
- NoiseScheduleManipulator: Adjusts noise schedules per block
- FeatureMapBlender: Caches and blends intermediate features
- SamplerInterceptor: Wraps the sampler for injection
These methods represent different strategies for achieving similar effects to block-level activation editing. They're not currently integrated but provide alternative approaches if the main method has limitations.
wanvideo_activation_patch.py
Example code showing how to manually patch WanVideoWrapper to support activation editing. This demonstrates:
- Where to add activation support in the model code
- How to read configuration from transformer_options
- The blending logic for context injection
This is useful if you want to:
- Understand how the activation system would work natively
- Manually patch WanVideoWrapper yourself
- Implement similar functionality in other models
Status
These files are not imported or used by the main nodes. They're kept for:
- Reference and documentation
- Future development possibilities
- Understanding alternative approaches
- Educational purposes