search

xuekt98 / BBDM

BBDM: Image-to-image Translation with Brownian Bridge Diffusion Models
MIT License
273 stars 28 forks source link

Generated Output Images information needed. #23

Open Nischit290402 opened 1 year ago

Nischit290402 commented 1 year ago

Hi, Can I know how to get the output image of the paired translation? In my results/.../LBBDM-f4/0/Image/train_sample directory, I have three images: Condition, ground_truth and skip_sample after training. Are we supposed to use the skip sample to obtain the desired image, or is skip_sample itself the generated image? I would really appreciate any help I can get since I am lost with this. P.S. : I am just trying to reimplement using the celeb dataset with masks in A and ground_truth in B. The paper does say to use VQGAN to get the final image IA->B. But I couldn't locate this specifically in the codebase. Thank You in advance.

xuekt98 commented 1 year ago

skip_sample is the generated image. :)

QiFengSu commented 12 months ago

skip_sample is the generated image. :)

3@M}P@~@GW TD9{D_35@436 Hello, thank you very much for you and your team for making the code publicly available. I currently have a question and hope to get your assistance. My generated skip_sample looks like this, and I have no clue about it. I'm not sure if you have the time to help me understand it. I would greatly appreciate it!

yug125lk commented 11 months ago

skip_sample is the generated image. :)

3@M}P@~@GW TD9{D_35@436 Hello, thank you very much for you and your team for making the code publicly available. I currently have a question and hope to get your assistance. My generated skip_sample looks like this, and I have no clue about it. I'm not sure if you have the time to help me understand it. I would greatly appreciate it!

Hi, did you solve it?

barkindagda commented 11 months ago

Hi, I am having the same results for the generated image after 10 epochs. Did anyone manage to figure out what is the reason for this ?

Thanks

collinsrakowski commented 11 months ago

I had the same issue

forhollow commented 11 months ago

To obtain accurate skip samples results, it's essential to ensure that the pre-trained model weight files are matched exactly (ckpt_path). For example, when using the ckpt_path for LBBDM-f16.yaml in VQGAN, I used the pre-trained weight file for vq-f16 provided by the author through the given link.

barkindagda commented 11 months ago

Thank you for your answer @forhollow . I am unable to access the pretrained weights provided by the authors because of the Baidu (It does not accept UK mobile number).I tried training the model with random initialisation and still got same results. I am using Night2day dataset. Do you have any advice ?

Thank you.