About the detail of Masked attention control

[siqi0905]

In equation (9), how is I obtained? Is the dimension of I different in each attention layer? How are they obtained to generate 𝑀𝑎𝑡𝑡𝑛 [𝑖, 𝑗]? Thank you very much for your response！

[ysy31415]

Hi, since we have the binary mask for input image, so we know the indices of known/unknown pixels. And yes, for layers with different resolutions (e.g., 64,32,16,8) , I has different sizes, the mask should be resized to match the corresponding resolution.

发件人: 万思琪 @.> 发送时间: 2023年10月14日 19:14 收件人: ysy31415/unipaint @.> 抄送: Subscribed @.***> 主题: [Ext] [ysy31415/unipaint] About the detail of Masked attention control (Issue #2)

CAUTION: External email. Do not reply, click on links or open attachments unless you recognize the sender and know the content is safe.

In equation (9), how is I obtained? Is the dimension of I different in each attention layer? How are they obtained to generate 𝑀𝑎𝑡𝑡𝑛 [𝑖, 𝑗]? Thank you very much for your response！

— Reply to this email directly, view it on GitHubhttps://github.com/ysy31415/unipaint/issues/2, or unsubscribehttps://github.com/notifications/unsubscribe-auth/BBVALB3NQXZNKT7F6XCYIODX7JXYHANCNFSM6AAAAAA6AF5KTA. You are receiving this because you are subscribed to this thread.Message ID: @.**@.>>

[siqi0905]

if the size of the binary mask for the input image is (1, 512, 512), do you mean we just need to resize it to (1, 64, 64), (1, 32, 32), ..., (1, 8, 8)?

[ysy31415]

Yes, that’s right