MBCNN trained on UHDM

Two inference-only MBCNN checkpoints reproduced with the standalone PyTorch implementation in Running-Turtle1/demoireing-resources.

File	Epoch	Selection	Full UHDM PSNR	Full UHDM SSIM
`uhdm/best_epoch110/model.safetensors`	110	Best PSNR on 100 center crops	20.223574	0.771880
`uhdm/last_epoch150/model.safetensors`	150	Fixed final epoch	20.002245	0.768795

The epoch-110 model is test-selected because the official UHDM test set was used for center-crop checkpoint selection. The epoch-150 model is provided as the fixed-schedule result. Both full evaluations use all 500 test images, direct FP32 full-resolution inference without tiling, ESDNet-style padding, full-RGB PSNR, and valid-window SSIM. LPIPS was not computed.

The files contain only model parameters in safetensors format. Training optimizer and scaler states are intentionally omitted. See the resource repository for code, exact configuration, per-image metrics, licensing, and attribution.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support