AI Paper Atlas

Research map / Image generation / Autoregressive and latent generation

Perceptual Artifacts Localization for Image Synthesis Tasks

Lingzhi Zhang, Zhengjie Xu, Connelly Barnes, Yuqian Zhou, Qing Liu, He Zhang, Sohrab Amirghodsi, Zhe Lin, Eli Shechtman, Jianbo Shi

ICCV 2023 · Image generation › Autoregressive and latent generation · approach: Generalizable fake image detection · 28 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
train one unified artifact segmentation model on all tasks instead of per-task specialists (ablation)Swin-T + UperNet artifact detector · 10 synthesis tasksmIoU: PSR 21.33 -> 5.10; StyleGAN2 37.85 -> 38.53; AnyRes 9.15 -> 34.74Mixed
fine-tune the pretrained detector with about 10 labeled images on an unseen generator (ablation)artifact detector · Versatile DiffusionmIoU: 5.92 -> 22.18Improved
fine-tune the pretrained detector with about 10 labeled images on an unseen generator (ablation)artifact detector · StyleGAN3mIoU: 46.45 -> 40.81Hurt
fine-tune the pretrained detector with about 10 labeled images on an unseen generator (ablation)artifact detector · DiTmIoU: 16.46 -> 31.76Improved
train on the new multi-task artifact dataset versus prior detectorsartifact segmentation · StyleGAN3 (unseen)mIoU: 11.43 -> 46.45Improved
use PAR as no-reference quality metric versus MUSIQ and HyperIQAimage quality ranking · Stable Diffusion pairsagreement with human preference: 48.45% -> 63.92%Improved
refine detected artifact regions with zoom-in inpainting using SD inpaintingStable Diffusion text2image · 5,000 SD imagesFID-CLIP: 15.98 -> 13.28Improved
gradually zoom in the inpainting input context (ablation)diffusion inpainting of faces and hands · 100 imagesPerceptual Artifacts RatioImproved
use CNNgenerates real/fake classifier with Grad-CAM as artifact localizerexisting detector baseline · 10 synthesis tasksmIoU: StyleGAN2 4.38 vs 38.53 for unified modelHurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy