Realism Control One-step Diffusion for Real-World Image Super-Resolution (RCOD)

Official pretrained model checkpoints for RCOD (AAAI 2026).

• GitHub Repository: https://github.com/Zongliang-Wu/RCOD
• Project Page: https://zongliang-wu.github.io/RCOD-SR/

Checkpoints Included

File	Base Architecture	Description
rcod_o.pkl	OSEDiff (SD 2.1)	Unified RCOD-O denoiser with LDG (Fidelity / Neutral / Realism)
rcod_s.pkl	S3Diff (SD-Turbo)	Unified RCOD-S denoiser with LDG (Fidelity / Neutral / Realism)
rcod_s_mem.pkl	MLP	Metric Estimation Module (MEM) for adaptive realism selection

External Dependencies

For training and external components, please download the respective third-party models from their original authors:

• DAPE (DAPE.pth): Domain-Adaptive Prior Extractor from OSEDiff / DAPE.
• RAM (ram_swin_large_14m.pth): Recognize Anything Model from xinyu1205/recognize-anything.
• de_net (de_net.pth): Degradation Estimator from S3Diff (included in RCOD_S/assets/mm-realsr/).

Quick Download

# Using huggingface-cli
pip install huggingface_hub
huggingface-cli download MMQDD/RCOD --local-dir weights/