Encountered some difficulties in reproducing

[zjumsj]

Dear Author,

I would like to express my sincere admiration and gratitude for your outstanding work. The relighting results from the pretrained head model of subject AXE977 are impressive. I would like to know the specific settings under which you trained this pretrained model.

I noticed that the pretrained model comes with a config.yml file, and based on its contents, I speculate that the parameters you used might be batch=4 and iters=600,000. However, I am having difficulty reproducing the results and am unsure where I might have gone wrong. Here are my training details:

GPU: One A800
CUDA Version: 11.8
Batch Size: 4
Iterations: 600,000

Additionally, I noticed many warnings like below. Is that expected behavior?

Thank you for your assistance.

[una-dinosauria]

The logger errors for missing data are expected. Unfortunately I am unsure of what went wrong with the rest of the training. These kinds of artifacts aren't expected at all, and I have recently re-trained all the models myself with different results (as you show).

Does this error reproduce if you train the model again? It could be a bad run with poor convergence -- these happens once in a while.

[zjumsj]

Thank you for your reply! I haven't made multiple attempts, as training takes several days. I am going to run it a few more times as suggested to see if the issue persists. Thank you again for your assistance, and for your outstanding work!

[una-dinosauria]

The quality seems very degraded, you may consider training for 100k iterations and see if these artifacts are still present (they shouldn't be).

At 100k things should look a bit blurred, but not too far off the final result otherwise.