qhjqhj00 vbench-t2v
Generative Refinement Networks for Visual Synthesis — Jian Han et al. (arXiv:2604.13030, 2026)
What this evaluates
Evaluates text-to-video generation quality and semantic alignment across dimensions like human action, scene composition, object consistency, and aesthetic quality.
Datasets
- VBench — total ?; splits: test (-1)
Metrics
VBench Overall(primary) — range: percent- Composite score aggregating Human Action, Scene, Multiple Objects, Appearance Style, Quality Score, and Semantic Score. Reported as percentage.
Input / output format
Input: Text prompt describing video content
Output: Video clip (2-10 seconds, 256x256 to 480p)
Scoring recipe
videos = []
for prompt in vbench_prompts:
videos.append(model.generate(prompt))
scores = vbench.evaluate(videos)
overall = mean(scores.dimensions)
Common pitfalls
- Scores are computed on videos generated after prompt rewriting.
- Automated VBench metrics may not fully capture temporal coherence or human preference.
Evidence (verbatim from paper)
GRN exhibits superior performance in generating videos from textual prompts. When benchmarked against contemporary diffusion and flow-based models—including AnimateDiff-V2 [animatediff], VideoCraft-2.0 [videocrafter], OpenSora V1.2 [opensora], Show-1 [show-1], and CogVideoX-5B [cogvideox]—GRN achieves significantly higher scores across quality, semantic, and overall scores.
Citation
@misc{han2026generativerefinement,
title={Generative Refinement Networks for Visual Synthesis},
author={Jian Han et al.},
year={2026},
note={arXiv:2604.13030}
}
- arXiv: 2604.13030