Official Comment · ICLR 2026

# Response to Reviewer – Weaknesses & Questions We thank the reviewer for pointing out the importance of qualitative comparisons and generalization demonstrations. We address each comment below. ## 1. On the lack of visual comparisons in the demo videos We agree that qualitative video results are crucial for evaluating Talking Face Synthesis (TFS). While the main paper focuses on quantitative metrics for space reasons, we have prepared extensive visual comparisons against multiple state-of-the-art approaches in the supplementary materials. Specifically, the file **“2-visual_comparative_results.pptx”** included in Supplementary_Material_Reply.zip provides side-by-side video frame comparisons between MuEx and strong baselines, including recent multilingual and cross-lingual TFS systems. These comparisons clearly show improvements in lip–audio synchrony, temporal stability, and facial realism. We will highlight this more explicitly in the revised manuscript and can also integrate selected comparative frames in the main paper if space permits. ## 2. On generalization to unseen image/person styles We appreciate the reviewer’s concern regarding out-of-domain generalization. The current paper focuses on multilingual audio generalization, but we agree that demonstrating robustness to unseen identities and styles is also important. To address this, we provide a dedicated set of experiments in the supplementary file: **“1-generalization_to_unseen_person&styles.pptx”**, which includes: - Unseen identities not present in training, - Different portrait styles, lighting conditions, and facial structures, - Cross-identity transfer tests, where MuEx drives faces with attributes far outside the training distribution. These results show that MuEx maintains consistent lip-sync accuracy and temporal coherence even when the input face significantly diverges from the training domain, supporting the model’s adaptability. ## Summary - We provide extensive visual comparisons with state-of-the-art methods in **2-visual_comparative_results.pptx**. - We demonstrate generalization to unseen identities and styles in **1-generalization_to_unseen_person&styles.pptx**. We thank the reviewer again for highlighting these valuable points, which we believe will further improve the clarity and completeness of our work.