CLEAR: Context-Aware Learning with End-to-End Mask-Free Inference for Adaptive Video Subtitle Removal

The two-stage training strategy is well-motivated: Stage I uses pixel-difference pseudo-labels with orthogonality constraints ($\mathcal{L}_{\text{ortho}}$) and adversarial purification to separate subtitle from content features without manual masks, while Stage II internalizes adaptive weighting into LoRA-augmented attention via the occlusion head $\mathcal{H}$. The parameter efficiency claim (0.77% trainable parameters) is specific and verifiable given the base Wan2.1-Fun 1.3B model and rank-64 LoRA configuration. Table 2's ablation demonstrates that each component (prior learning, context distillation, generation feedback) contributes measurably to the final performance, with the full system achieving 26.80 dB PSNR versus 21.62 dB for LoRA-only baseline.

The primary limitation is data availability and reproducibility: the 160,000 video pairs are custom-collected and not publicly available, making independent verification impossible. The pixel-difference pseudo-labeling (Equation 2: $\hat{\mathbf{M}}_{t}(i,j)=1$ if $\Delta_t > \mu_t + \sigma_t$) assumes subtitles are visually distinct from background, which fails for semi-transparent text, motion-blurred regions, or subtitles with colors matching the scene. While the paper emphasizes 'mask-free inference,' the method still requires paired training data (subtitled + clean) which is often harder to obtain than binary masks for existing inpainting methods. Additionally, the impressive zero-shot generalization to six languages (Figure 1) lacks quantitative metrics—Table 1 reports results only on Chinese subtitles, leaving open questions about performance degradation on non-Latin scripts or complex typography.

The comparison to baselines (ProPainter, MiniMax-Remover, DiffuEraser) appears fair: all methods use identical binary masks generated via the same thresholding procedure, ensuring the performance gains stem from CLEAR's adaptive generation mechanism rather than better input masks. The evaluation protocol is comprehensive, spanning reconstruction (PSNR, SSIM), perceptual (LPIPS, DISTS, VFID), and temporal consistency metrics (TWE, TC). However, the baseline comparison conflates architectural differences with input modality differences—CLEAR receives the full subtitled video while baselines receive explicit masks, so the comparison actually demonstrates that internal adaptive weighting outperforms external binary masking, not necessarily that CLEAR outperforms these architectures when all use the same inputs. The dramatic VFID reduction (-74.7% vs. best baseline) suggests strong perceptual quality, but without human evaluation or user studies, the subjective 'clean removal' claims remain unvalidated.

The authors provide code availability via GitHub, and Appendix B contains detailed hyperparameters: LoRA rank 64 applied to all attention and FFN layers, AdamW optimizer with learning rates $2\times 10^{-5}$ (Stage I) and $1\times 10^{-4}$ (Stage II), and specific resolution-adaptive training (1280×720). However, reproduction is severely hindered by the private dataset requirement (160,000 Chinese subtitle pairs). The paper uses Wan2.1-Fun-V1.1-1.3B as the base model, which is publicly available, but training requires 8 GPUs and approximately 1 day for Stage II convergence. Critical missing details include: exact data synthesis pipeline for the 100 green-screen pairs, specific font distributions in the training set, and whether the 400 test samples represent a held-out set from the same distribution or distinct content. The occlusion head architecture (2.1M parameters) is specified as two convolutional layers, but the exact layer dimensions and activation patterns beyond Equation 13 are not provided.