cnulab / RealNet

Offical implementation of "RealNet: A Feature Selection Network with Realistic Synthetic Anomaly for Anomaly Detection (CVPR 2024)"
MIT License
183 stars 9 forks source link

Generating Anomaly Synthesis Images of Custom Sizes #48

Closed MinGiSa closed 1 month ago

MinGiSa commented 1 month ago

Thank you for sharing your good work. By the way, is there a way to generate anomaly synthesis images in sizes other than 256x256, such as 512x512 or 512x1024?

cnulab commented 1 month ago

Hello! We used the pre-trained diffusion model provided by BeatGans for anomaly synthesis, so the resolution of the synthesis depends on BeatGans. If you want to train a 512*512 diffusion model, you need to download the weights from BeatGans and then modify the input_size and unet(classifier) structure in experiments/MVTec-AD/diffusion.yaml and experiments/MVTec-AD/classifier.yaml . The unet(classifier) structure should refer to the settings in BeatGans.

MinGiSa commented 1 month ago

@cnulab thank you !