search

Zhendong-Wang / Prompt-Diffusion

Official PyTorch implementation of the paper "In-Context Learning Unlocked for Diffusion Models"
Apache License 2.0
364 stars 10 forks source link

train #10

Open Tanghui2000 opened 7 months ago

Tanghui2000 commented 7 months ago

Hello, thank you for your work, which is enlightening to me. I have some questions to consult. I tried to train grayscale images and color images as a set of "examples", so as to achieve image coloring. However, the generated images are very different from the original ones. Why does this happen? How to ensure consistency with the original image?

Zhendong-Wang commented 7 months ago

The consistency is maintained because the model is also given the query image as the inputs. As we show in the Figure 3. The training also depends on the dataset, you input and output image pairs should be consistent.

Tanghui2000 commented 7 months ago

The consistency is maintained because the model is also given the query image as the inputs. As we show in the Figure 3. The training also depends on the dataset, you input and output image pairs should be consistent.

Thank you for helping me. It is now possible to keep the input consistent with the content of the generated image. But the resulting image's contrast and saturation seem to have changed, making it seem more noisy than the original, which bothers me. I'm using the coco dataset. Some of the images in the coco dataset are noisy. Is this misleading? Or is it something else? 屏幕截图 2023-11-29 091525

Zhendong-Wang commented 7 months ago

Yes, here I think image quality is important for finetuning. If the dataset is noisy, you may need longer training to converge and may converge to worse results.