AI图片超分2026：Real-ESRGAN免费把480p变4K

🔧 AI工具 2026-08-02 约 5 分钟阅读

老照片、压缩截图和低分辨率产品图放大后惨不忍睹。AI超分是重建细节，不是拉伸像素。

Real-ESRGAN: The Free Upscaler That Surprised Everyone

Real-ESRGAN (Real-ESRGAN: Practical Algorithms for General Image/Video Restoration, 36,351 GitHub stars, BSD-3-Clause) is the open-source image upscaler that became the default answer for restoring and enlarging images. It was trained on real-world degradations - compression artifacts, blur, noise - rather than clean synthetic downscales, which is why it produces natural detail instead of oversharpened mush.

What It Does Better Than Bicubic

Bicubic interpolation just smooths pixels; the result is soft and blurry at 4x. Real-ESRGAN generates plausible detail - edges, textures, fine structure - using a generative model. A 480p screenshot upscaled 4x with Real-ESRGAN looks like a native high-res image for most practical purposes: product photos, old family pictures, game captures, document scans.

How to Use It

# pip install realesrgan (or use the WebUI)
python inference_realesrgan.py -n RealESRGAN_x4plus -i input.jpg -o output/

Or use the bundled GUI (Windows) / integrated versions in Upscayl, a popular free desktop app that wraps Real-ESRGAN (and other models) in a one-click interface.

The Model Family

RealESRGAN_x4plus - the general-purpose 4x model.
RealESRGAN_x4plus_anime_6B - tuned for anime/manga; sharper linework without halos.
Real-ESRGAN-ncnn-vulkan - Vulkan build that runs on CPU and weak GPUs.
RealESRGAN_x2 - 2x upscale, less hallucinated detail, good for slight enlargements.

Honest Limitations

Faces and text can get weird at aggressive scales - generative models invent detail that may not match reality. For archival restoration of important images, upscale at 2x and combine with manual cleanup. For anything where authenticity matters (evidence, documents), keep the original and treat the upscale as a display copy.