image generation quality

[xLuge]

hi, nice work, but , the image reconstruction quality is very low, What details need to be noted during inference?

[HelenMao]

Thank you for your interest in our work! We apologize for the bugs in the code. The problem has now been fixed, and the updated version is available.
Here are the results after running the corrected code:

[Looperswag]

The input for inference should be a 4x4 grid with 1024x1024 resolution?

[DreamLoveBetty]

The input for inference should be a 4x4 grid with 1024x1024 resolution?

[Orannue]

Hi @Looperswag @DreamLoveBetty , thank you for your attention！ If you use our provided model weights, we highly recommend using a 4×4 grid with a 1024×1024 resolution as input, since our dataset follows this format.

We have not tested data in other formats yet. However, you can try using inputs with different resolutions in a 4×4 grid and modify the mask resolution in the following code:

Edit-Transfer/edit_transfer.py

Line 27 in c44a846

mask=torch.zeros(1,1024,1024)

Alternatively, you can construct a new dataset in a different format and test it.

Let us know if you have any questions!

[Looperswag]

Hi @Looperswag @DreamLoveBetty , thank you for your attention！ If you use our provided model weights, we highly recommend using a 4×4 grid with a 1024×1024 resolution as input, since our dataset follows this format.

We have not tested data in other formats yet. However, you can try using inputs with different resolutions in a 4×4 grid and modify the mask resolution in the following code:

Edit-Transfer/edit_transfer.py

Line 27 in c44a846

mask=torch.zeros(1,1024,1024)

Alternatively, you can construct a new dataset in a different format and test it.
Let us know if you have any questions!

Is there any quick way to construct a 4*4 grid given images?

[Orannue]

Hi @Looperswag @DreamLoveBetty , thank you for your attention！ If you use our provided model weights, we highly recommend using a 4×4 grid with a 1024×1024 resolution as input, since our dataset follows this format.
We have not tested data in other formats yet. However, you can try using inputs with different resolutions in a 4×4 grid and modify the mask resolution in the following code:
Edit-Transfer/edit_transfer.py
Line 27 in c44a846
mask=torch.zeros(1,1024,1024)
Alternatively, you can construct a new dataset in a different format and test it.
Let us know if you have any questions!

Is there any quick way to construct a 4*4 grid given images?

The quickest approach is using FLUX to directly generate a 4×4 grid image, but the results are often unsatisfactory.
So we generate a 1×2 grid first and then manually select two results to compose a 4×4 grid. Alternatively, we use off-the-shelf editing methods to modify images and construct the final 4×4 grid.