Unable to Reproduce Paper Results with Provided Pre-trained Models

[aliciusschroeder]

Hi, thank you for releasing your code and pre-trained models. I've been trying to reproduce the results from your paper on FFHQ and CelebA-HQ but am observing significant discrepancies from the reported performance. I wanted to document my findings in case there's something I'm missing or a known issue.

Environment & Setup

• Repository commit: eb4e77c84d00678e409343b52d9804b8ff31f467 (retrieved 2026-01-14)
• Pre-trained models: Iteration 700,000 checkpoints downloaded from BaiduCloud (as linked in README)
• Test data: FFHQ and CelebA-HQ with mixed contamination types (scribbles + rectangular image patches)

Observed Issues

I've attached a comparison image showing Input → GT → Predicted Mask → Output → Predicted Output columns across 4 test samples.

1. Complete Failure on Rectangular Occlusions

The model fails entirely on square/rectangular image patches (rows 2 & 4 in attached image):

• The calendar text obstruction remains fully visible in the output
• The predicted mask localizes only a small central region rather than the full occlusion
• The 'Output column shows extreme noise (scattered white pixels across the entire image), suggesting the mask estimation has collapsed

This contrasts with the paper's Table 3 results showing strong performance on "Image Occlusion" contamination patterns.

2. Poor Texture Quality on Successful Detections

Even where the model correctly identifies contamination (scribble rows 1 & 3):

• Inpainted regions exhibit over-smoothing / "plastic skin" artifacts
• Loss of high-frequency detail (pores, skin texture) compared to GT
• Color blending issues leaving visible discoloration blobs (especially around mouth/chin areas)

3. Facial Feature Reconstruction

• Lips appear undefined and blurry when occluded
• Eye reconstruction shows asymmetry and lack of definition

Questions

1. Are the BaiduCloud checkpoints the same ones used to generate the paper's quantitative results?
2. Is there specific preprocessing required for the test images beyond resizing to 256×256?
3. Were the paper results obtained with a different contamination synthesis pipeline than what's in the released code?
4. Any known issues with certain contamination types (solid rectangles vs. irregular masks)?

Attached

• comparison.jpg: Side-by-side comparison showing the issues described above

I'd appreciate any guidance on reproducing the reported results. Happy to provide additional details or test specific configurations if helpful.

Thanks for your time!

[Haoru]

On the complete failure with square / rectangular masks

This issue indeed stems from an oversight during our code release. In the initial version of the repository, the released pre-trained model was not fully consistent with the corresponding data generation code, which led to abnormal behavior when the model encountered regular square or rectangular occlusions.

We would like to explicitly emphasize that during training, our model was exposed to square and rectangular masks. Specifically, we adopted a free-form mask generation strategy, which includes not only irregular scribble-like masks but also block-shaped square and rectangular occlusions.

However, in the initial public release, the mismatch between the data generation code and the provided pre-trained model caused the model’s mask prediction to degrade on regular rectangular occlusions, giving the impression that the model had not been trained on such masks. This mismatch ultimately resulted in the observed restoration failures.

We sincerely apologize for the confusion. We have now fixed the dataset generation code and updated the corresponding pre-trained model to ensure consistency with the training setting described in the paper. We warmly welcome you to reproduce the results with the updated version and share your observations.

On the over-smoothing of texture in successfully detected cases

It is important to emphasize that under the Unseen Contaminated Patterns setting (Table 3), the contaminations used at test time correspond to degradation patterns that are not observed during training, i.e., the model is evaluated under unseen contamination distributions. Our work follows a typical blind contamination removal paradigm.

Under this setting, the model is required to simultaneously localize and remove contaminated regions without any prior knowledge of the contamination type or location. While some successfully detected cases may still exhibit relatively smooth textures, the model is able to accurately identify the corrupted regions and effectively remove the interference, demonstrating strong robustness compared to existing methods, particularly under unseen contamination conditions.

It should be noted that directly comparing restoration results under such cross-domain, unseen contamination settings with models trained and evaluated on the same contamination distribution, and expecting identical visual quality, is inherently challenging—especially for high-frequency texture reconstruction (e.g., fine facial details). Improving texture fidelity under such blind settings remains an important direction for future work.