Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

AI Upscaling Needs Provenance Labels for Reliable Image Workflows

AI summary

A Juejin AI article compares interpolation, CNN, and GAN super-resolution methods on three artificially degraded 1024×1024 samples and finds that model-generated pixels can introduce or remove image detail, especially on portraits and text. It recommends recording the method, scale, model and weight version, and whether the original remains available, while retaining source assets for comparison, pricing, and archival workflows. The author also proposes downscaling and PSNR checks as a fallback for detecting unmarked upscaled images.

In-depth insight

The article gives asset teams a concrete way to distinguish generated pixels from camera-captured data and avoid treating upscaled images as originals in verification or archival systems.

一条图片入库的链路上挂个超分环节是很自然的想法。缩略图糊了就先放大再切,列表页立刻好看。麻烦在后头:放大完的那张图一旦进了库就和用户当初传上来的那张混在一块了。再往后没人分得出哪个像素是相机给的、哪个是模型补的。下游只是列表页的封面图的话混了也没事。下游要做比对、核价和存档的时候混进去的就是事故。

这一轮的数是这么来的。三张 1024×1024 的样张先用 Pillow 11.3.0 的 LANCZOS 缩到 256 或 512,再让各种办法放回 1024 跟原图逐像素比,4× 那一组是八种办法、2× 那一组少一种。插值那边是 Pillow 的 nearest、bicubic 和 Lanczos。经典 CNN 超分走 OpenCV 5.0.0 的 dnn_superres 配官方预训练权重 ESPCN、FSRCNN、EDSR 和 LapSRN。GAN 那一类的数我用的是图映 ImgIng imging.cn/ 工作台的第 1 档快速高清,面板上显示跑的是 Real-ESRGAN General x4v3,五档里只跑了这一档。SSIM 用的是 scikit-image 0.24.0,PSNR 和分块统计是我自己写的。机器是一台 Apple 芯片 Mac 配桌面 Chromium,每种办法各跑一次。降质是人造的这件事得先交代:样张本来就清楚,是缩小以后再放回去的。真实的糊照片还叠着对焦不准、手抖、高 ISO 噪点和转发攒下来的 JPEG 损伤,这套数字搬不到那种图上去。

#Real-ESRGAN General x4v3#ImgIng#Pillow#OpenCV#scikit-image#ESPCN#FSRCNN#EDSR#LapSRN

Original · 7 min readRead the original
Group chat

No comments yet. Start the conversation.