AuraSR V2, the latest version of the single-step GAN upscaler, has been released by the team, addressing several limitations identified in its predecessor. While AuraSR V1 was based on the Adobe Gigagan paper and faced issues upscaling JPG compressed images without artifacts, the new version incorporates a degradation process similar to ESRGAN training to better handle various image degradations. Improvements in V2 also include a revised training approach using 256-pixel tiles of 1024-pixel images, which aligns more closely with the inference process and helps reduce excessive detail addition. Another enhancement is the introduction of the upscale_4x_overlapped inference method, which minimizes seams in images by averaging results from overlapping tiles. Although AuraSR V2 retains the same architecture as V1, making it a straightforward replacement, it better preserves details in "in the wild" images compared to other models like RealESRGAN, and the model is now available on Huggingface and deployed to fal's AuraSR endpoint. Plans for AuraSR V3 include using higher resolution images and a new architecture to further enhance the model's capabilities.