Poor Output Quality Using demo_sample.ipynb with Provided Model Parameters

[yuanzhang7]

Thank you for providing the pretrained model parameters and the demo_sample.ipynb notebook. However, I encountered an issue where the generated results are of very poor quality and do not match the input text at all. Could you please help me understand and resolve this issue? Here is an example image with text of 'an apple':

[daixiangzi]

First, you need to make a more detailed description, such as there is a red apple on the table; in addition, the category you need to describe should be relatively common in the imagenet dataset. I think apples may not be particularly common in this dataset. You can try replacing apples with cars.

[Dreamern]

I have the same problem. Here is an example image with text of "there is a car on the road"

[daixiangzi]

I have the same problem. Here is an example image with text of "there is a car on the road"

ok，I will check my checkpoint weight which is right?

[Ankowa]

I have the same problem with the default prompt in the demo_sample.ipynb "'a sailboat with a sail at sunset'"

[Sologala]

Same poor results

[daixiangzi]

Text: "a sailboat with a sail at sunset"

[daixiangzi]

I have the same problem with the default prompt in the demo_sample.ipynb "'a sailboat with a sail at sunset'"

Sorry, I found that I uploaded the wrong weight. I am updating the latest weight.

[daixiangzi]

Text:'a photo of a car on the road'

[daixiangzi]

@Ankowa @Dreamern @Sologala @yuanzhang7

[daixiangzi]

Text:'a photo of an apple on the table'

[daixiangzi]

Another thing to note is that the text description can be more and the description method can be adjusted more, because these will greatly affect the generated effect. After all, we just trained it once on ImageNet. If we want better results, I think we need more detailed text and diverse image data.

[daixiangzi]

last-ckpt have updated. you can try it again.