AuthFace: Towards Authentic Blind Face Restoration with Face-oriented Generative Diffusion Prior

1. In Figure 2(b), the fine-tuned model shows only marginal improvements over the baseline SDXL model, which brings into question whether the fine-tuning process effectively enhances facial detail restoration.
2. While the study emphasizes the use of photographic prompts during fine-tuning, it lacks specific experiments to verify if these prompts genuinely contribute to improved model performance.
3. The authors did not incorporate photographic prompts in the second stage of training or during testing. Exploring whether adding these prompts could further enhance the model’s restoration quality would be beneficial.
4. Since AuthFace uses a specially collected, high-quality face dataset, the comparison with other methods might not be entirely fair. Competing methods didn’t have access to this dataset, which could make it hard to tell if AuthFace’s improvements come from the method itself or just from having better data.

My concerns are as follows:

1. About the proposed dataset. After a careful reading, I observe that the proposed dataset is only used in stage one, i.e., used to fine-tune the prestrained T2I model, which means that the dataset seems to be solving the problem of generating faces only, not for the BFR task. This can also be seen in Figure 4, where the authors are using an example from the FFHQ dataset to plot.
2. Concerns about methodological innovativeness. The methodological contribution of this work is simply to fine-tune the existing T2I model, i.e., SDXL and ControlNet, and the novelty may be inadequate.
3. For the proposed time-aware latent facial feature loss, why consider only the eyes and mouth? For the facial structures, the nose is very important.

1. While using high-quality face images to fine-tune the pre-trained model is likely important for generating better face priors, the proposed method seems to rely on an additional 1,500 high-quality images, which may be only a trick compared to methods like BFRfusion and SUPIR.
2. In Fig. 6 and Fig. 7, the restored face images exhibit artifacts. For instance, the left eye is closed in the ground truth (GT) image, but in the restoration produced by AuthFace, the eye appears partially open.
3. Some universal evaluations should be conducted. The authors do not use the pre-trained T2I models (fine-tuned after incorporating the 1,500 high-quality images) on existing methods, which could validate the effectiveness of the proposed approach in comparison to those methods.
4. The proposed method seems very heavy. Some complexity results should be provided, such as parameters, FLOPs, and runtime. They can help to more fairly assess the true effectiveness of the method.

1. The visual image in Figure 1 does not convince me that Authface performs better than SUPIR in the task of missing details and incorrect details. For example, in the red box in the first row, Authface lacks more details than SUPIR.
2. In qualitative experiments, Authface does not perform as well as BFRffusion. For example, in the fourth row of Figure 6, the eyes of BFRffusion are closer to GT than Aythface, and in the first two rows of Figure 7, there is little difference between the two methods. In fact, the image quality of the teeth of BFRffusion is better.
3. More Sota compared methods. Please show more comparisons on more Sota methods, like DiffBIR and StableSR, which also leverage ControlNet as a part of the framework.
4. Contribution. As mentioned above, DiffBIR, StableSR, and so on, have already enhanced the generative capabilities of pretrained T2I models for BFR, which is not a novel research direction.