encode_jpeg generates noise when processing 4k image

[Lily-Git-hub]

Hi I tried the latest torchvision 19.0 with pytorch2.4. I found that the encode_jpeg func had a problem when processing 4k image. For example, I have a 4K image tensor, which size is (3, 2160, 3840), then I use torchvision.io.encode_jpeg() in a loop, in the first loop, it can encode the tensor correctly. But in the following loops, it will generate jpeg image of noise only. Could you help with this please? Thanks!

[NicolasHug]

Hi @Lily-Git-hub can you please provide a minimal reproducing example? Thank you

[Lily-Git-hub]

Hi Nicolas,

Please try this example:

import torch 
import torchvision
import torch.nn.functional as F   

for i in range(2):
    image_data = torch.load('image_data.pt')
    resized_image_tensor = F.interpolate(image_data.unsqueeze(0), size=(2160, 3820), mode='bilinear', align_corners=False)
    image_data_resized = resized_image_tensor[0]
    image_data_encoded = torchvision.io.encode_jpeg( (image_data_resized).to(torch.uint8) )
    data = image_data_encoded.cpu().numpy().tobytes()
    with open(f'1.jpg', 'wb') as f:
        f.write(data)

    del data, image_data_encoded, resized_image_tensor, image_data_resized, image_data`

without the last line of code, which deleted the used variables, the saved image would be noise only. Please unzip the 'image_data.zip' to get image_data.pt.
[image_data.zip](https://github.com/user-attachments/files/16656971/image_data.zip)

[NicolasHug]

Sorry @Lily-Git-hub , I cannot reproduce your issue.

[Lily-Git-hub]

del data, image_data_encoded, resized_image_tensor, image_data_resized, image_data`

Hi Nicolas,

Did you remove the above line of code？ The error occurs when not deleting used variables. Thanks!

[NicolasHug]

Yes, I deleted these lines. Can you please provide a more minimal reproducing example, without a for loop, wihtout resizing, and from a normal image rather than from a pt file (which I won't load on my machine for security reaosns)

[glazhh]

I encountered a similar issue. I resolved it by adding torch.cuda.synchronize() before using encode_jpeg. It seems there might be some synchronization problems between F.interpolate and torchvision.io.encode_jpeg.

   resized_image_tensor = F.interpolate(image_data.unsqueeze(0), size=(2160, 3820), mode='bilinear', align_corners=False)
   image_data_resized = resized_image_tensor[0]
   # add synchronize after modified image and before encode jpeg
   torch.cuda.synchronize()
   image_data_encoded = torchvision.io.encode_jpeg( (image_data_resized).to(torch.uint8) )