search

va1shn9v / PromptIR

PromptIR: Prompting for All-in-One Blind Image Restoration [NeurIPS 2023]
https://arxiv.org/abs/2306.13090
Other
297 stars 22 forks source link

A question about the prompt in the paper. #24

Closed shallweiwei closed 3 months ago

shallweiwei commented 5 months ago

A great work! I'm confused about the prompt-component.It seems to be fixed across tasks. But I'm starting to think that it should be different between tasks, is there a problem with my understanding, so how does the model differentiate between tasks?

https://github.com/va1shn9v/PromptIR/blob/644e529287035ade5824f816f2caaf81883cbfb8/net/model.py#L221

Yaziwel commented 5 months ago

A great work! I'm confused about the prompt-component.It seems to be fixed across tasks. But I'm starting to think that it should be different between tasks, is there a problem with my understanding, so how does the model differentiate between tasks?

https://github.com/va1shn9v/PromptIR/blob/644e529287035ade5824f816f2caaf81883cbfb8/net/model.py#L221

The feature maps of inputs from various tasks will predict varying weights to assign to N prompts, thereby generating the final prompt. Throughout this process, prompt generation evolves in response to the input.

shallweiwei commented 5 months ago

So the various of prompts are because of the feature maps？It's kind of like a classification task. Is there any possibility of misclassification, such as denoising tasks being classified as deraining tasks?