Yueming6568
Some questions about the implementation of open source code.
Thank you so much for open source. I have a few questions about some details:
1.I noticed that in your inference.py file, you use a file called fs3.npy. Could you explain what the data in this file means and how I can obtain it? I couldn't find any code in your repository for generating this file.
2.I also noticed that in your train.py file, you don't set a max_iters parameter, and only iterate through the dataset once. Does this mean that the delta mapper doesn't need multiple rounds of training?
3.Your open source code doesn't seem to include the code for cspace_ffhq_feat.npy and sspace_ffhq_feat.npy. Could you provide this code or tell me where I can find it?
4.When I tried editing the sample images you provided with the default parameters, I got the same results as shown in the paper for the provided texts. However, when I tried other texts, the results weren't very good. But when I used the text "Blue hair", I got a result that was close to the "Blue suit" in the paper. Have you encountered this situation before?
I also cannot find the file fs3.npy. Could you please tell me how you ran the program and obtained the result? Thank you.
You can find this file in the styleclip project and hope it can help you.
Thank you for your questions! Regarding Q1, the fs3.npy file contains matrix Rs which records how CLIP image embeddings change when modifying each dimension in S space - please refer to the disentanglement section in our paper. For Q2, we actually trained for 50 epochs (code in ./datasets/train_dataset.py). The files mentioned in Q3 are available on Google Drive: https://drive.google.com/drive/folders/13NLq4giSgdcMVkYQIiPj4Xhxz4-wlXSD. About Q4, DeltaSpace's semantic expressiveness is constrained by CLIP's training data distribution, particularly for rare attributes (e.g., "blue hair") or novel attribute combinations absent in training data. We plan to address these limitations through domain-adaptive fine-tuning in future work.
