search

ali-vilab / FlashFace

MIT License
387 stars 37 forks source link

training code #7

Open isidentical opened 7 months ago

isidentical commented 7 months ago

Any plans to release the training code? would love to see this trained over a realistic model (like realistic vision or dreamshaper) or better adapted to XL!

jshilong commented 7 months ago

Any plans to release the training code? would love to see this trained over a realistic model (like realistic vision or dreamshaper) or better adapted to XL!

The training code heavily relies on the dataset we constructed, which can be referred to in our report. Releasing the training code along with the dataset poses a significant challenge since the dataset is sensitive and requires careful handling within the company.

isidentical commented 7 months ago

Would love if you guys released it without the dataset. Even a non working version can save a lot of time and heaches for people who want to reproduce the work!!

On Mon, Apr 22, 2024 at 19:10 Shilong Zhang @.***> wrote:

Any plans to release the training code? would love to see this trained over a realistic model (like realistic vision or dreamshaper) or better adapted to XL!

The training code heavily relies on the dataset we constructed, which can be referred to in our report. Releasing the training code along with the dataset poses a significant challenge since the dataset is sensitive and requires careful handling within the company.

— Reply to this email directly, view it on GitHub https://github.com/ali-vilab/FlashFace/issues/7#issuecomment-2071279497, or unsubscribe https://github.com/notifications/unsubscribe-auth/ALJKHQJSWXM7DNCMJ6Z5VY3Y6W7JHAVCNFSM6AAAAABGTZI4DGVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDANZRGI3TSNBZG4 . You are receiving this because you authored the thread.Message ID: @.***>

jshilong commented 7 months ago

Since my internship at Alibaba has ended and I have resigned, I am unable to personally pursue this matter. I will contact my colleagues to inquire about the feasibility of this matter.

On the other hand, the core code is flashface/all_finetune/models/reference_unet.py and has excellent portability. I believe you can quickly adapt it to the existing open-source SD training code. You can also give it a try.