Overall visual quality captures how clearly edges are resolved and how naturally fine textures are rendered. In these YouHQ examples, DOVE retains the main scene structure but softens fine details such as masonry and roof tiles. RelayVSR renders these textures more distinctly, which can make the video look clearer and more natural.
Participants judged overall quality, fidelity to the visible LR content, and temporal stability separately. Smoother outputs can retain the structure visible in LR and make small texture changes less noticeable during playback. Table 11 reports GSB scores for RelayVSR against DOVE of +53.6% in overall quality, −8.9% in content fidelity, and −8.7% in temporal stability. The YouHQ examples illustrate how spatial detail can contribute to the overall preference.
Capture the selected update window. With zoom on, click a paused video to choose the same region in all views.
What to watch
The reviewer asks how post-training changes perceptual quality, fidelity, and continuity at reference updates. The videos compare Full VARO with No post-training, Reference-only, and System-only. Step through the same region before and after an update to see whether texture changes abruptly or carries through to the next non-keyframes.
R1-Table2 reports PSNR and full-video warping error alongside LPIPS, CLIP-IQA, and MANIQA. Full VARO scores best on all three perceptual metrics outside the training reward. In the blinded comparisons in R1-Table3, viewers favored Full VARO for smoother reference updates over No post-training and Reference-only. Against System-only, the larger preference margin was for overall quality.
A reference frame can be sharp even when the intermediate frames look softer. Fast camera or subject motion moves details across the image, while occlusion exposes regions that the endpoint references may poorly represent. Compare a moving subject or a newly visible region at a reference frame and again between references to see where detail enhancement weakens.
A heavily degraded LR input leaves fine detail ambiguous. Neighboring generated references can reconstruct different textures for the same region or introduce hallucinated details. The lightweight VSR network can carry these errors into intermediate frames as artifacts. When the reference changes, the texture can shift or flicker, making the output less temporally stable. Open LR + GT and step across a reference update to check the affected region against the ground truth.
A hard cut can fall between fixed-interval references, placing the two endpoints in different shots. Compare the last frame of the outgoing shot and the first few frames of the incoming shot. The red markers on the timeline indicate the cuts. The LR preview is blurred and resampled for display. Both model outputs use the original LR.
RelayVSR uses features from the current frame to query the keyframe references and video memory. This lets it assign less weight to references from a different shot. During training, randomly sampled clips can span shot boundaries.