search

KaiChen1998 / GeoDiffusion

Official PyTorch implementation of GeoDiffusion in ICLR 2024 (https://arxiv.org/abs/2306.04607)
https://kaichen1998.github.io/projects/geodiffusion/
MIT License
61 stars 3 forks source link

Generation quality of the nuimages-model #11

Closed zhangxiao696 closed 4 months ago

zhangxiao696 commented 4 months ago

hello, when I follow your codes train nuimages dataset, the clarity of image generation is very low n008-2018-09-12-09-06-08-0400__CAM_BACK_LEFT__1536758322547405_2

n008-2018-09-14-14-02-31-0400__CAM_BACK__1536948703937558_4

KaiChen1998 commented 4 months ago

@zhangxiao696 Actually, I would say these demo images demonstrate quite good quality :), considering the default setting we provide is based on 256x256 resolution, which is significantly lower than the original 1600x900 real images.

I do recommend you increase the image resolution to 512x512 or even larger. What you need to do is simply modify the image resolution in the data config file and scale your learning rate linearly with you batch size. Consider adopting the gradient accumulation via the --gradient_accumulation_steps argument if necessary for your GPU resources.