search

lllyasviel / LayerDiffuse_DiffusersCLI

LayerDiffuse in pure diffusers without any GUI
Apache License 2.0
248 stars 18 forks source link

About Dataset #5

Open XinYu-Andy opened 2 weeks ago

XinYu-Andy commented 2 weeks ago

Hi, Thanks for your great work! I am wondering if you would release the dataset in your paper?

In fact, i am trying to reproduce the dataset creation pipeline, but I found it was very hard to inpaint the foreground object (i.e, remove this object without generating new ones!). Do you have some suggestions? Like what specific SDXL version do you use and if there exists some magic prompt tricks..

Thank you in advance!

lllyasviel commented 2 weeks ago

Hi,

Thanks for the message. The dataset contains purchased items with a complex mixture of licenses, so we cannot provide it directly. However, I will consider other methods in the future to make it more accessible.

Regarding inpainting without adding any object, beyond the paper version from months ago, I found some other effective methods in recent weeks:

  1. Use LAMA inpainting for a coarse inpaint and then use SD inpainting with a denoise setting of around 0.6 to 0.7.
  2. Alternatively, you can purchase the Adobe inpaint API and pay for processing each image in the dataset if your dataset is not too large.
  3. Try using the positive prompt "an image of nothing" and the negative prompt "lots of objects" with SDXL at CFG 7.5.
XinYu-Andy commented 2 weeks ago

wow, thanks for your quick response. Using LAMA for a coarse inpaint and then using SD I2I for refinement is exactly what I am doing today, what a coincidence! But I haven’t found it useful yet. I will try more examples and adjust the prompt you suggested.

Thanks!

On Thu, 27 Jun 2024 at 22:59, lllyasviel @.***> wrote:

Hi,

Thanks for the message. The dataset contains purchased items with a complex mixture of licenses, so we cannot provide it directly. However, I will consider other methods in the future to make it more accessible.

Regarding inpainting without adding any object, beyond the paper version from months ago, I found some other effective methods in recent weeks:

  1. Use LAMA https://github.com/advimman/lama inpainting for a coarse inpaint and then use SD inpainting with a denoise setting of around 0.6 to 0.7.
  2. Alternatively, you can purchase the Adobe inpaint API and pay for processing each image in the dataset if your dataset is not too large.
  3. Try using the positive prompt "an image of nothing" and the negative prompt "lots of objects" with SDXL at CFG 7.5.

— Reply to this email directly, view it on GitHub https://github.com/lllyasviel/LayerDiffuse_DiffusersCLI/issues/5#issuecomment-2194952434, or unsubscribe https://github.com/notifications/unsubscribe-auth/AOPNNM3UTZ4PEMSTKVL2Z5LZJQSFDAVCNFSM6AAAAABJ76SII2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDCOJUHE2TENBTGQ . You are receiving this because you authored the thread.Message ID: @.***>

XinYu-Andy commented 2 weeks ago

Hi,

Thanks for the message. The dataset contains purchased items with a complex mixture of licenses, so we cannot provide it directly. However, I will consider other methods in the future to make it more accessible.

Regarding inpainting without adding any object, beyond the paper version from months ago, I found some other effective methods in recent weeks:

  1. Use LAMA inpainting for a coarse inpaint and then use SD inpainting with a denoise setting of around 0.6 to 0.7.
  2. Alternatively, you can purchase the Adobe inpaint API and pay for processing each image in the dataset if your dataset is not too large.
  3. Try using the positive prompt "an image of nothing" and the negative prompt "lots of objects" with SDXL at CFG 7.5.

BTW, is the background conditioned layer diffusion already available in Comfyui (I am not familiar with this tool so I want to ask this question)? Since it has not been available yet in your diffuser-only repo.