Unified Benchmark for Semantic-Payload Watermarking

Construct a unified empirical benchmark that directly compares the caption-based real-valued semantic watermarking pipeline of SWIFT with the binary CLIP-VAE semantic watermarking framework under comparable backbone and evaluation conditions.

Background

The paper discusses SWIFT as the closest concurrent work to CLIP-VAE. SWIFT embeds image captions as high-dimensional real-valued vectors using a modified HiDDeN backbone, whereas CLIP-VAE encodes CLIP image embeddings as 100-bit binary watermarks using the VINE backbone and channel-aware bit-flip training.

A direct comparison is not provided because the two systems use different captioning pipelines, watermarking backbones, and payload representations. Establishing a unified benchmark would enable a controlled assessment of binary versus real-valued semantic payloads and clarify the trade-off between semantic capacity and robustness to noisy watermark channels.

References

We do not provide a direct empirical comparison since SWIFT's caption-based pipeline requires a separate captioner and HiDDeN backbone that differ from our setup; constructing a unified benchmark is left for future work.

— Semantic Watermarking for Malicious Image Manipulation Detection  (2609.39623 - Kim et al., 30 Sep 2026) in Section 2, Related Work, paragraph “Concurrent Semantic-Payload Watermarking”

Continuous-valued watermarks (e.g., SWIFT) allow higher capacity but sacrifice binary-channel robustness; a unified benchmark of binary vs. real-valued semantic payloads is left for future work.

— Semantic Watermarking for Malicious Image Manipulation Detection  (2609.39623 - Kim et al., 30 Sep 2026) in Section 5, Limitations, paragraph “Other Untested Alternatives and Concurrent Work”