- Rating
- 1(−1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper presents SEDiT, a mask-free video subtitle removal framework based on LTX-Video. It combines reference-token conditioning, LoRA fine-tuning, subtitle-region loss weighting, and one-step inference. The authors introduce VSR-Bench-400, a synthetic paired benchmark, and an optional distillation stage using real failure cases. Experiments cover reconstruction quality, inference speed, and minute-scale video processing.
Critical strengths
- The paper addresses a practical problem. Removing test-time subtitle localization can simplify processing for stylized or difficult-to-segment subtitles.
- The reported quality–speed trade-off is promising. Region-resolved metrics, a VAE reconstruction reference, and chunk-boundary experiments provide useful supporting evidence.
Critical weaknesses
-
The comparisons confound reconstruction quality with input information. In Table 1, mask-based baselines lose all content inside the subtitle box, while SEDiT retains the background between characters and information beneath translucent text. Appendix G.6 acknowledges that stroke-level scoring does not remove this asymmetry. Quantitative comparisons with precise masks and composited baseline outputs are needed to establish how much of the advantage comes from the proposed model rather than the input protocol.
-
The methodological novelty and the explanatory value of the theory are limited. Mask-free removal and sequence-concatenated conditioning have precedents acknowledged in the paper. The main method combines these with standard fine-tuning and loss weighting. The conditional-mean result is valid, but neither conditional uncertainty nor endpoint field error is measured, so it does not explain the trained model's one-step performance. A matched direct conditional-reconstruction baseline would clarify the contribution of flow matching.
-
Real-video usability and the distillation gains are not sufficiently established. The 30-video usability criterion checks only subtitle residuals, allowing background damage or temporal artifacts to go unpenalized. The improvement from 25/30 to 29/30 therefore establishes more complete erasure, not overall usability. The distillation data size and additional training budget are also unspecified, and there is no matched continued-training or random-mining control to isolate the benefit of failure-driven distillation.
Questions for authors
- How do the results change when baselines receive precise subtitle masks and preserve unedited pixels through compositing?
- Under matched initialization, data, and training budgets, how does one-step flow matching compare with direct conditional latent reconstruction?
- Can you jointly evaluate residuals, background damage, and temporal consistency on real videos, and isolate failure-driven distillation from additional data and training?