Train on my own dataset, and the generated images are all images from my dataset, with strong repeatability

[ZZZBBBZZZ]

Medfusion is an amazing work.Thanks for your work! I tried to generate on my own blastocyst image, which is a small dataset with only 440 samples. The biggest problem currently encountered is:

1. When using to generate a blastocyst image, the generated blastocyst image is almost identical to the image in my dataset, except for some differences in shadows. They cannot be considered as new images.
2. There are a large number of duplicate images in the generated images. To take an extreme example, the generated 100 images may all be one image in the dataset with different shadows added. Is it because some of my parameters have not been adjusted properly?Why haven't new images appeared? Can you give me some advice and help?

[Ystartff]

Hello, did you solve this problem? I also have a high image repetition rate in my experiments, where an image may be generated multiple times

[ZZZBBBZZZ]

Hello, did you solve this problem? I also have a high image repetition rate in my experiments, where an image may be generated multiple times

Sorry, I haven't solved this problem yet.

[Shame-fight]

Medfusion is an amazing work.Thanks for your work! I tried to generate on my own blastocyst image, which is a small dataset with only 440 samples. The biggest problem currently encountered is:

1. When using to generate a blastocyst image, the generated blastocyst image is almost identical to the image in my dataset, except for some differences in shadows. They cannot be considered as new images.
2. There are a large number of duplicate images in the generated images. To take an extreme example, the generated 100 images may all be one image in the dataset with different shadows added. Is it because some of my parameters have not been adjusted properly?Why haven't new images appeared? Can you give me some advice and help?

Hello, my dataset is also very small, with approximately 600 images. But the results generated by training are very poor, with significant differences from the original image. Can you tell me the complete process from training to outputting and generating images, and which files need to be modified? I'm not sure where I did poorly。By the way, are you Chinese?

[ZZZBBBZZZ]

Medfusion is an amazing work.Thanks for your work! I tried to generate on my own blastocyst image, which is a small dataset with only 440 samples. The biggest problem currently encountered is:

1. When using to generate a blastocyst image, the generated blastocyst image is almost identical to the image in my dataset, except for some differences in shadows. They cannot be considered as new images.
2. There are a large number of duplicate images in the generated images. To take an extreme example, the generated 100 images may all be one image in the dataset with different shadows added. Is it because some of my parameters have not been adjusted properly?Why haven't new images appeared? Can you give me some advice and help?

Hello, my dataset is also very small, with approximately 600 images. But the results generated by training are very poor, with significant differences from the original image. Can you tell me the complete process from training to outputting and generating images, and which files need to be modified? I'm not sure where I did poorly。By the way, are you Chinese?

Hello. My training process primarily follows the steps outlined in README.md. I was unable to generate ideal images previously, so I attempted to reduce the image resolution(512512->256256). This may be of help to you. Finally, I am a Chinese.

[Shame-fight]

Medfusion is an amazing work.Thanks for your work! I tried to generate on my own blastocyst image, which is a small dataset with only 440 samples. The biggest problem currently encountered is:

1. When using to generate a blastocyst image, the generated blastocyst image is almost identical to the image in my dataset, except for some differences in shadows. They cannot be considered as new images.
2. There are a large number of duplicate images in the generated images. To take an extreme example, the generated 100 images may all be one image in the dataset with different shadows added. Is it because some of my parameters have not been adjusted properly?Why haven't new images appeared? Can you give me some advice and help?

Hello, my dataset is also very small, with approximately 600 images. But the results generated by training are very poor, with significant differences from the original image. Can you tell me the complete process from training to outputting and generating images, and which files need to be modified? I'm not sure where I did poorly。By the way, are you Chinese?

Hello. My training process primarily follows the steps outlined in README.md. I was unable to generate ideal images previously, so I attempted to reduce the image resolution(512_512->256_256). This may be of help to you. Finally, I am a Chinese.

谢谢，我训练VAE模型不到100个epoch就停止了，得到的指标数值还可以。但训练完diffusion模型后，运行sample_dataset. py生成的图像非常差，没有学习到任何东西。我是用的是开源甲状腺超声图像DDTI数据集，约600张图像，分辨率统一为 512*512.所以我觉得可能我哪些参数没有修改正确。并且得到的precision数值为0。方便的话是否能加一下你的微信，像你请教一下。或者是否能提供在你数据集上修改后的代码，我对比看一下。

[ZZZBBBZZZ]

Medfusion is an amazing work.Thanks for your work! I tried to generate on my own blastocyst image, which is a small dataset with only 440 samples. The biggest problem currently encountered is:

1. When using to generate a blastocyst image, the generated blastocyst image is almost identical to the image in my dataset, except for some differences in shadows. They cannot be considered as new images.
2. There are a large number of duplicate images in the generated images. To take an extreme example, the generated 100 images may all be one image in the dataset with different shadows added. Is it because some of my parameters have not been adjusted properly?Why haven't new images appeared? Can you give me some advice and help?

Hello, my dataset is also very small, with approximately 600 images. But the results generated by training are very poor, with significant differences from the original image. Can you tell me the complete process from training to outputting and generating images, and which files need to be modified? I'm not sure where I did poorly。By the way, are you Chinese?

Hello. My training process primarily follows the steps outlined in README.md. I was unable to generate ideal images previously, so I attempted to reduce the image resolution(512_512->256_256). This may be of help to you. Finally, I am a Chinese.

谢谢，我训练VAE模型不到100个epoch就停止了，得到的指标数值还可以。但训练完diffusion模型后，运行sample_dataset. py生成的图像非常差，没有学习到任何东西。我是用的是开源甲状腺超声图像DDTI数据集，约600张图像，分辨率统一为 512*512.所以我觉得可能我哪些参数没有修改正确。并且得到的precision数值为0。方便的话是否能加一下你的微信，像你请教一下。或者是否能提供在你数据集上修改后的代码，我对比看一下。

欢迎交流！我的的邮箱是zheng_bozhong@163.com,我们可以通过邮箱来交换微信。

[taranrai]

Medfusion is an amazing work.Thanks for your work! I tried to generate on my own blastocyst image, which is a small dataset with only 440 samples. The biggest problem currently encountered is:

1. When using to generate a blastocyst image, the generated blastocyst image is almost identical to the image in my dataset, except for some differences in shadows. They cannot be considered as new images.
2. There are a large number of duplicate images in the generated images. To take an extreme example, the generated 100 images may all be one image in the dataset with different shadows added. Is it because some of my parameters have not been adjusted properly?Why haven't new images appeared? Can you give me some advice and help?

Do you have the evaluation metrics for these results too? I'm wondering if the precision and recall measures are able to sufficiently tell us whether the generated outputs are really similar to the real images.

[ZZZBBBZZZ]

Because the images I generated did not visually meet my expectations, I did not pay much attention to precision and recall rates, and they were vary low every time.

Medfusion is an amazing work.Thanks for your work! I tried to generate on my own blastocyst image, which is a small dataset with only 440 samples. The biggest problem currently encountered is:

1. When using to generate a blastocyst image, the generated blastocyst image is almost identical to the image in my dataset, except for some differences in shadows. They cannot be considered as new images.
2. There are a large number of duplicate images in the generated images. To take an extreme example, the generated 100 images may all be one image in the dataset with different shadows added. Is it because some of my parameters have not been adjusted properly?Why haven't new images appeared? Can you give me some advice and help?

Do you have the evaluation metrics for these results too? I'm wondering if the precision and recall measures are able to sufficiently tell us whether the generated outputs are really similar to the real images.

Because the images I generated did not visually meet my expectations, I did not pay much attention to precision and recall rates, and they were vary vary low every time.

[taranrai]

Many thanks for the fast reply. Interesting to know. I'm particularly interested in the histopathology use cases, where from your medfusion paper, the diffusion models do not appear to have too low recall and precision values, especially when compared to the cGAN. Nonetheless, I'm wondering if it is the domain being problematic itself, for example, extracting patches from histological slides could contribute towards limiting diversity as many of the images are already almost "identical", almost repeating, like wallpaper.

[clearlyzero]

Medfusion is an amazing work.Thanks for your work! I tried to generate on my own blastocyst image, which is a small dataset with only 440 samples. The biggest problem currently encountered is:

1. When using to generate a blastocyst image, the generated blastocyst image is almost identical to the image in my dataset, except for some differences in shadows. They cannot be considered as new images.
2. There are a large number of duplicate images in the generated images. To take an extreme example, the generated 100 images may all be one image in the dataset with different shadows added. Is it because some of my parameters have not been adjusted properly?Why haven't new images appeared? Can you give me some advice and help?

Hello, my dataset is also very small, with approximately 600 images. But the results generated by training are very poor, with significant differences from the original image. Can you tell me the complete process from training to outputting and generating images, and which files need to be modified? I'm not sure where I did poorly。By the way, are you Chinese?

Hello. My training process primarily follows the steps outlined in README.md. I was unable to generate ideal images previously, so I attempted to reduce the image resolution(512_512->256_256). This may be of help to you. Finally, I am a Chinese.

谢谢，我训练VAE模型不到100个epoch就停止了，得到的指标数值还可以。但训练完diffusion模型后，运行sample_dataset. py生成的图像非常差，没有学习到任何东西。我是用的是开源甲状腺超声图像DDTI数据集，约600张图像，分辨率统一为 512*512.所以我觉得可能我哪些参数没有修改正确。并且得到的precision数值为0。方便的话是否能加一下你的微信，像你请教一下。或者是否能提供在你数据集上修改后的代码，我对比看一下。

欢迎交流！我的的邮箱是zheng_bozhong@163.com,我们可以通过邮箱来交换微信。 I tried to implement it simply using pytorch according to the author's idea. I found that it could not be generated and the model could not converge.