2xSPAN_ch48_clean

2× SPAN (Swift Parameter-free Attention Network) super-resolution, feature_channels=48, trained clean on DIV2K (no degradation/JPEG-compensation). Repackaged from the official hongyuanyu/SPAN release (spanx2_ch48.pth), Apache-2.0.

params wrapper stripped → flat state_dict saved as .safetensors for direct loading.

Why clean (not degradation-trained): for upscaling clean VAE-decoded diffusion frames, degradation-trained SPANs (multijpg/ldl) over-smooth real texture by "correcting" compression artifacts that aren't present. The clean DIV2K variant matches-or-beats compact-pretrain sharpness.