Question about image resolution and upscaling in AIDT Dataset

[ThreeSRR]

Hi, this is an impressive work!
I noticed that the AIDT dataset uses InsightFace Swapper to generate synthetic images. Since the swapping model runs at 128×128 resolution, did you upscale the results using standard interpolation (like Lanczos/Bicubic) or an AI-based method such as CodeFormer?

[dsj320]

i think the author use the interpolation method,

[deepcs233]

Yes, we used standard interpolation to upscale the results from the InsightFace Swapper, which operates at a 128×128 resolution. We did not apply AI-based super-resolution methods like CodeFormer. This choice was made because the decoupled face is only used for feature extraction, so high-fidelity identity preservation and portrait quality are not required in this stage.

[ThreeSRR]

Thanks a lot for your helpful response!