Enhanced Latent-Space Adversarial Training for Super-Resolution

Real-world super-resolution (SR) at large upscaling factors (i.e., ≥ 4×) remains difficult due to complex real-image degradations. HYPIR, a leading diffusion-based restoration model, performs strongly on many inputs, yet for a non-trivial portion of more challenging cases a single forward step does not fully recover fine-grained details. A naive two-stage cascade improves visual quality, but introduces over-saturation, weak texture details, and high inference latency. To address these issues, this paper proposes HYPIR++. It removes the degradation removal encoder and noise augmentation modules to better preserve fidelity cues from low-quality inputs. Equipped with an optimized latent ConvNeXt and a latent patch discriminator, HYPIR++ supports latentspace adversarial learning for clearer details and more stable local structures. It further shortens text sequences and replaces full attention with sparse neighbor attention, enabling efficient highresolution inference without tiling. Experiments show that HYPIR++ improves perceptual quality and runs 1.71× faster than HYPIR on large-factor real-world SR

Have you ever tried to enlarge a small, blurry photo and ended up with something that looks worse than the original? Restoring badly damaged images at large magnifications (4× or more) is a long-standing challenge: the algorithm must remove noise and blur while also inventing fine details that aren't physically there. Today's leading method, HYPIR, handles many photos in a single pass but struggles on the hardest ones; running it twice in a row helps, yet introduces oversaturated colors, fuzzy textures, and painfully slow processing on large images. We introduce HYPIR++, which fixes all three issues at once. It strips out internal modules that erase the input photo's true color information, performs its "realism training" in a compressed feature space so it concentrates on the details people actually see, and replaces costly global attention with a "neighbor-only" variant that processes big images in one shot instead of stitching tiles together. The result is noticeably sharper, more natural restorations delivered 1.71× faster than before. This brings practical, large-factor photo restoration—useful for old family photos, low-resolution surveillance footage, and scientific imaging—much closer to everyday use.