The reproduction results were poor

[3Dian1415926-xin]

black_point---ground_truth ,,purple_point---predict_point ,,red,green,blue--fusion_point

I wrote an fusion code code myself ，shanghai_part_A MAE:240 ！ Do not use written fusion code ,MAE:270

[CharlieSong1999]

Hi,
Could you please share your reproduction code? The code in this repo seems incomplete and has loads of debugging code (like the "assert False" in the training script). Moreover, I can't find the "counting branch".

Thanks for any help.

[3Dian1415926-xin]

I gave up reproducing this code. If you want it to run successfully, you can refer to the original diffusion model. The author's code is rather messy. However, I still suggest choosing a different direction. I spent two months on it and found that the experimental results deviated significantly from the paper's indicators. I hope you can achieve good results!