DeepFillv2_Pytorch icon indicating copy to clipboard operation
DeepFillv2_Pytorch copied to clipboard

Published 20 hours ago • verified publisher icon csqiangwen

Trouble with quality of results using pretrained model

Open dwallener opened this issue 3 years ago • 11 comments

Used the pretrained model against a test image of my own, and the results are not good at all. Curious if there are any thoughts on what might be causing this. Everything seems to run correctly.

03-with-cutout 03-mask 5_second_out (1)

dwallener avatar Jun 20 '21 23:06 dwallener

It is a little weird. This case should be easily handled by this inpainting algorithm.

Have you tried another image?

csqiangwen avatar Jul 31 '21 11:07 csqiangwen

Uploading image.png… I also hava a same problem,use your pretrained model,the result was bad.And I used your code to traing celeba_hq datasets,the results also wsa bad.Why is it happen?

wmmin avatar Sep 25 '21 06:09 wmmin

Excuse me. My training samples are also poor in quality. (trained for about 125K iterations) Do I need to train it longer? Or the model itself is mistaken? epoch6batch0_gt epoch6batch0_second_out P.S. The 1st pic is ground truth, the 2nd pic is the result. (Also it was in 256 resolution but the paper said it can reach 512)

AlonzoLeeeooo avatar Oct 11 '21 12:10 AlonzoLeeeooo

me too

wmnj avatar May 09 '22 11:05 wmnj

me too

Tiuuuuu avatar Jan 29 '24 13:01 Tiuuuuu

Excuse me. My training samples are also poor in quality. (trained for about 125K iterations) Do I need to train it longer? Or the model itself is mistaken? epoch6batch0_gt epoch6batch0_second_out P.S. The 1st pic is ground truth, the 2nd pic is the result. (Also it was in 256 resolution but the paper said it can reach 512)

I have tried another PyTorch code implementation of DeepFill-v2 using mmcv and it works. You may refer to it if you are interested.

AlonzoLeeeooo avatar Jan 30 '24 05:01 AlonzoLeeeooo

Excuse me. My training samples are also poor in quality. (trained for about 125K iterations) Do I need to train it longer? Or the model itself is mistaken? epoch6batch0_gt epoch6batch0_second_out P.S. The 1st pic is ground truth, the 2nd pic is the result. (Also it was in 256 resolution but the paper said it can reach 512)

I have tried another PyTorch code implementation of DeepFill-v2 using mmcv and it works. You may refer to it if you are interested. How did you train DeepFill using mmcv？

poinne avatar Mar 12 '24 07:03 poinne

Excuse me. My training samples are also poor in quality. (trained for about 125K iterations) Do I need to train it longer? Or the model itself is mistaken? epoch6batch0_gt epoch6batch0_second_out P.S. The 1st pic is ground truth, the 2nd pic is the result. (Also it was in 256 resolution but the paper said it can reach 512)

I have tried another PyTorch code implementation of DeepFill-v2 using mmcv and it works. You may refer to it if you are interested. How did you train DeepFill using mmcv？

125k iterations might not be sufficient to reach the optimal performance of DeepFill-v2. I recommend training DeepFill-v2 with 300k iterations or above.

AlonzoLeeeooo avatar Mar 12 '24 07:03 AlonzoLeeeooo

Excuse me. My training samples are also poor in quality. (trained for about 125K iterations) Do I need to train it longer? Or the model itself is mistaken? epoch6batch0_gt epoch6batch0_second_out P.S. The 1st pic is ground truth, the 2nd pic is the result. (Also it was in 256 resolution but the paper said it can reach 512)

I have tried another PyTorch code implementation of DeepFill-v2 using mmcv and it works. You may refer to it if you are interested. How did you train DeepFill using mmcv？

125k iterations might not be sufficient to reach the optimal performance of DeepFill-v2. I recommend training DeepFill-v2 with 300k iterations or above.

Nonono, I'm curious about how you implemented DeepFill using mmcv. Did you rewrite the network structure of DeepFill using components from mmcv? If possible, could you provide me with a reference code?

poinne avatar Mar 15 '24 09:03 poinne

Excuse me. My training samples are also poor in quality. (trained for about 125K iterations) Do I need to train it longer? Or the model itself is mistaken? epoch6batch0_gt epoch6batch0_second_out P.S. The 1st pic is ground truth, the 2nd pic is the result. (Also it was in 256 resolution but the paper said it can reach 512)

I have tried another PyTorch code implementation of DeepFill-v2 using mmcv and it works. You may refer to it if you are interested. How did you train DeepFill using mmcv？

125k iterations might not be sufficient to reach the optimal performance of DeepFill-v2. I recommend training DeepFill-v2 with 300k iterations or above.

Nonono, I'm curious about how you implemented DeepFill using mmcv. Did you rewrite the network structure of DeepFill using components from mmcv? If possible, could you provide me with a reference code?

MMCV has an official code. You can directly use that implementation. Details can be referred to this website.

AlonzoLeeeooo avatar Mar 15 '24 09:03 AlonzoLeeeooo

Excuse me. My training samples are also poor in quality. (trained for about 125K iterations) Do I need to train it longer? Or the model itself is mistaken? epoch6batch0_gt epoch6batch0_second_out P.S. The 1st pic is ground truth, the 2nd pic is the result. (Also it was in 256 resolution but the paper said it can reach 512)

I have tried another PyTorch code implementation of DeepFill-v2 using mmcv and it works. You may refer to it if you are interested. How did you train DeepFill using mmcv？

125k iterations might not be sufficient to reach the optimal performance of DeepFill-v2. I recommend training DeepFill-v2 with 300k iterations or above.

Nonono, I'm curious about how you implemented DeepFill using mmcv. Did you rewrite the network structure of DeepFill using components from mmcv? If possible, could you provide me with a reference code?

MMCV has an official code. You can directly use that implementation. Details can be referred to this website.

Sure, thank you very much!

poinne avatar Mar 15 '24 09:03 poinne