Questions about reproduction

[DragonKingA]

Thank you for your excellent work, which is crucial to advancing the field. However, I encountered some issues during reproduction and would appreciate your clarification:

1. I attempted to reproduce the results (batchsize=8, patchsize=8, iteration=400k), but even using the provided checkpoint, I couldn't achieve metrics close to those reported in the paper (LPIPS: 0.456, FID: 22.047, ArtFID: 33.564, CFSD: 0.246). I've made multiple attempts without success. Could there be an issue?
2. I also reproduced MambaST concurrently. When testing both MambaST and SaMam on the same content and style images, I found that MambaST consistently achieved much lower FID and ArtFID than SaMam—a trend contrary to the supplemental material provided.
3. My measurements of MACs and inference time yielded 28.57 GB and 69.76 ms on average, which differ from the values in the paper. Is the code complete, and what might be causing this discrepancy in my reproduction?

[Chernobyllight]

Thank you for your excellent work, which is crucial to advancing the field. However, I encountered some issues during reproduction and would appreciate your clarification:

1. I attempted to reproduce the results (batchsize=8, patchsize=8, iteration=400k), but even using the provided checkpoint, I couldn't achieve metrics close to those reported in the paper (LPIPS: 0.456, FID: 22.047, ArtFID: 33.564, CFSD: 0.246). I've made multiple attempts without success. Could there be an issue?
2. I also reproduced MambaST concurrently. When testing both MambaST and SaMam on the same content and style images, I found that MambaST consistently achieved much lower FID and ArtFID than SaMam—a trend contrary to the supplemental material provided.
3. My measurements of MACs and inference time yielded 28.57 GB and 69.76 ms on average, which differ from the values in the paper. Is the code complete, and what might be causing this discrepancy in my reproduction?

1. Our open source checkpoints are trained on content loss weight=7.0 and style loss weight=7.0. You can try to adapt them manually for better style metrics (e.g. bigger style loss weight).

2. For best of our knowledge, the checkpoint with patchsize=4 could achieve better quantitative metrics. However, its computational efficiency will be lower than setting patchsize=8.

3. Image resolution: our method processes content images on original scale, and can also specify style image resolution in test code. And MambaST processes content and style images on fixed resolution. This may also affect metrics.

[ygtreceee]

So in order to ensure that the indicators in the paper can be reproduced normally, can you provide the optimal performance configuration of the model? (e.g., loss weight, batch_size, patch_size, iteration, lr)