Reusing one sinusoidal block recurrently expands spectral support without extra parameters: by Jacobi–Anger, each sine step creates new frequency combinations.

We characterize the pre-softmax attention matrix QK in transformers as an associative memory matrix encoding pairwise associations between input features.

We introduce TAG, a theoretically grounded, training-free, computationally lightweight, and architecture-agnostic guidance method that operates solely on trajectory signals without modifying the underlying diffusion model.

We propose an image-based RAG framework (iRAG) for realistic super-resolution, using a trainable hashing function to retrieve either real-world or generated references from an LR query.

ReAL grounds diffusion SR in retrieved visual references rather than text priors. It encodes a matched reference once and reuses its cached attention features throughout denoising, enabling reference-driven texture and structure recovery without CFG.

We introduce a bit-plane decomposition that enables lossless implicit neural representations, allowing exact reconstruction of digital signals with continuous networks.