TRaM-VSR: Efficient One-Step Diffusion Video Super-Resolution via Token Routing
Researchers have introduced TRaM-VSR, a framework for Token Routing and Merging designed for one-step diffusion video super-resolution. This innovative approach tackles the quadratic computational demands associated with dense spatio-temporal token sequences in Diffusion Transformers (DiT). By integrating motion-sensitive temporal information with semantic text similarity, it assesses token significance, allowing for the differentiation of dynamic objects and structural boundaries. An offline planner adjusts the importance to facilitate routing through optimally clustered network blocks, ensuring that structurally vital tokens are maintained within each routed group. This technique seeks to prevent the irreversible loss of detail and temporal flickering, which are prevalent in efficiency-driven methods, particularly in one-step diffusion models. The paper can be accessed on arXiv.
Key facts
- TRaM-VSR is a Token Routing and Merging framework for one-step diffusion video super-resolution.
- It uses Diffusion Transformer (DiT) priors for exceptional perceptual quality.
- The method addresses quadratic computational cost of dense spatio-temporal token sequences.
- Token importance is estimated by fusing motion-sensitive temporal cues with semantic text similarity.
- An offline planner calibrates importance to guide routing across grouped network blocks.
- Structurally critical tokens are preserved within each routed group.
- The approach aims to avoid irreversible detail loss and temporal flickering.
- The paper is available on arXiv with identifier 2607.22231.
Entities
Institutions
- arXiv