Artscale (SuperResV13) is a 2× EDSR-style super-resolution network (256 channels, 64 residual channel-attention blocks, ICNR PixelShuffle head) with a fixed, phase-correct Lanczos-3 baseline added at the output. It was trained to restore lightly degraded art-site images (ArtStation / X / DeviantArt artifacts, mild under-sizing) to 2× without hallucinating texture. Three files are provided:
- The original (SuperResV13), trained on 100,000 source resolution image tiles for 80 epochs.
- The finetune (SuperResV13FT), trained on an additional 200,000 source resolution image tiles for an additional 10 epochs.
- The ComfyUI custom node zip, containing the necessary inference and tiling code.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support