Skip to content

Inquiry regarding dataset construction in TINKER #3

Description

@Jillianbug

Hello! Congratulations on your excellent work!

I am highly interested in the construction process of your multi-view consistent image editing dataset. To explore this, I recently conducted some experiments using the FLUX.1-Kontext-dev model, where I horizontally concatenated two images as the input.

However, I encountered a couple of issues during my experiments:

(1) Global Alteration during Local Editing: When the target editing area is strictly local, the model tends to rewrite the entire image, causing unintended changes to the background.

(2) Resolution and Sizing Inconsistency: The model seems to automatically resize my original images, which results in a mismatch between the dimensions of the input and the generated output.

I was wondering if you encountered these similar challenges when building your dataset? If so, how did you effectively resolve them?

I would truly appreciate your insights on this. I am eagerly looking forward to your reply, as well as the upcoming release of your code and datasets!

Thank you very much for your time and help.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions