search

CompVis / fm-boosting

FMBoost: Boosting Latent Diffusion with Flow Matching (ECCV 2024 Oral)
https://compvis.github.io/fm-boosting/
MIT License
183 stars 3 forks source link

Date about codes release #1

Closed ygean closed 1 month ago

ygean commented 10 months ago

when will you release codes? I read the paper, and it's a exciting work!

ziyizhetutanota commented 9 months ago

Same question !

benearnthof commented 9 months ago

Hello @zhouyuangan & @ziyizhetutanota, I've tried to produce a minimal working example here https://github.com/benearnthof/fm_boosting/tree/main Please feel free to check out the script there. I'm not too sure about the concatenation process that's depicted in Figure 3 of the paper so please raise issues or pull requests if you have suggestions. I used OT-CFM flow matching like it is implemented in TorchCFM https://github.com/atong01/conditional-flow-matching/tree/main

As of today this is just a minimal working example, training is slow but seems to work (I'll link samples tomorrow after 24hours of training on a A100 40GB GPU). I'll keep working on this and will wrap everything in a nice package in the coming days.

ygean commented 8 months ago

Hi. How was the training? Is there any update?

benearnthof commented 8 months ago

Hi @ygean, sorry for the delay. I've added noise concatenation with an image conditional UNet from imagen_pytorch by Phil Wang (Lucidrains on Github) and just started a training run on 128 -> 256 pixel images. Will take about two days. If you have any suggestions please feel free to raise an issue on my proof of concept implementation, so far results were blurry I'm hoping to get sharper images with the new UNet.

ygean commented 8 months ago

Hi @ygean, sorry for the delay. I've added noise concatenation with an image conditional UNet from imagen_pytorch by Phil Wang (Lucidrains on Github) and just started a training run on 128 -> 256 pixel images. Will take about two days. If you have any suggestions please feel free to raise an issue on my proof of concept implementation, so far results were blurry I'm hoping to get sharper images with the new UNet.

Hi，does the new unet work fine now?

benearnthof commented 8 months ago

Unfortunately not, my results are still blurry. I'll keep working on the implementation however as I want to integrate this model in a larger project of mine.

xiao2mo commented 3 months ago

when will you release codes?

m990130 commented 1 month ago

Hi we released the training code. In the next few days we will release the checkpoint and inference code.