Get Started
Home
Topics
Search
Library
Research questionHow can dominant-channel activations enable lightweight adaptation of frozen DiT-based super-resolution models?Improving perceptual quality in real-world super-resolution often requires fine-tuning the diffusion-transformer backbone or attaching additional adapters. That leaves the backbone’s concentrated activation structure underused as a potential adaptation interface.
Computer Vision
Diffusion Models
Image & Video Processing
Machine Learning
Latest papersRecent research connected to this question, newest first.SPARK: Input-Conditioned Sparse Activation Modulation for Frozen DiT-based Super-ResolutionThe source studies frozen DiT-based super-resolution backbones and a lightweight controller that modulates selected dominant channels using low-resolution VAE latents while keeping the backbone and VAE frozen. Evidence covers three backbones and the DIV2K, RealSR, and DRealSR datasets, with reported gains in fidelity and perceptual quality when eight channels per stream and block are modulated.research paper · Sep 3, 2026
Related questions
How can frozen dense-prediction models remain robust under distribution shift without retraining or changing their prediction heads?How can high-resolution diffusion Transformers prune tokens without sacrificing image fidelity or predictable compute?How can diffusion models preserve fine perceptual detail when reconstructing high-resolution images from low-resolution inputs?How does structured model compression change a model’s ability to adapt at test time under distribution shift?