search

Sherrylone / PQDiff

[ICLR 2024] Continuous-Multiple Image Outpainting in One-Step via Positional Query and A Diffusion-based Approach Link: https://arxiv.org/abs/2401.15652
68 stars 1 forks source link

Issues in generating an extended image #3

Closed tariqtomar closed 7 months ago

tariqtomar commented 8 months ago

I trained the model with the scenery dataset and when i am trying to generate an extended image it is creating a colorful noisy image. Not able to figure out what actually the issue is. I think there is something wrong in the decoder part. Please find the attached files for the reference. ori gen copy

Sherrylone commented 8 months ago

Could you provide the training log?

tariqtomar commented 8 months ago

Yeah sure output.log

Sherrylone commented 8 months ago

Hi, you should train the model with longer iterations, e.g., 100k steps.

tariqtomar commented 8 months ago

This are the training samples results. Prime anchor prime_anchor-1601

decode anchor decode_anchor-1601

prime target prime_target-1601

decode target decode_target-1601

predict target predict_target-1601

tariqtomar commented 8 months ago

actually i am testing with my custom dataset and I have only 150 samples to train, for that i have trained the model with 2000 steps and got similar results.

Sherrylone commented 8 months ago

Hi, although there are only 150 samples, I think 2000 steps are not enough to converge.

Therefore, perhaps you can change the model to a smaller size, and reduce the steps to 30k?

tariqtomar commented 7 months ago

Thanks for the reply. It worked with higher number of steps and by increasing some more samples.

ucasyjz commented 4 months ago

Thanks for the reply. It worked with higher number of steps and by increasing some more samples.

Can you give me the trained weight on Scenery, thank you