ABSTRACT
The paper introduces ACS-Gate, a novel approach to conditionally diffuse models that adapts the strength of conditioning based on the current timestep and denoising state, improving sample quality and fidelity.
PAPER · PDF
Loading PDF…
Key findings
ACS-Gate dynamically modulates conditioning strength based on timestep and denoising state.
A lightweight gating network predicts timestep-specific conditioning weights.
Theoretical analysis establishes optimality of timestep-adaptive conditioning.
The method is training-efficient with minimal computational overhead.
Limitations & open questions
The effectiveness of ACS-Gate is yet to be validated across diverse datasets and model architectures.
The paper does not discuss potential issues with overfitting to specific conditioning modalities.