search

locuslab / ect

Consistency Models Made Easy
166 stars 5 forks source link

ECT is a unconditional model? #1

Closed 2019211753 closed 2 months ago

2019211753 commented 2 months ago

I'm considering whether it's possible to modify the ECT model to generate samples based on text conditions. Additionally, is it possible to change the image mode to another mode? such as 3D human skeleton.

Gsunshine commented 2 months ago

Hi @2019211753 ,

Thanks for your interest in ECT! ECT can also work in conditional settings. We tested it using CIFAR10 and ImageNet. I think it works well for class conditional settings. Indeed, conditional information even boosts and eases learning, according to my observations.

You are welcome to try & test other modalities and conditioning. But do pay attention to monitor and cope with the "curse of consistency"!

Thanks, Zhengyang

2019211753 commented 2 months ago

Hi @Gsunshine , Thank you for your response! I have a few additional questions. Could you please provide guidance on how to set up the ECT environment? Also, how should I proceed to run it? I encountered an error when attempting to execute bash run_ecm.sh 3 12000 --desc bs128.200k: the error message was "run_ecm.sh: line 2: torchrun: command not found."

Furthermore, could you clarify if ECM and ECT refer to the same thing? Apologies if my lack of knowledge has led to too many questions.

Gsunshine commented 2 months ago

Hi @2019211753 ,

We've updated the environment setup. I suggest setting the number of GPUs as the power of 2. You need a GPU machine with at least 1 A6000 GPU or equivalent. Set up your environment using conda as in README.

Sorry for the confusion. We denote the technique from our blog post as Easy Consistency Tuning (ECT), while the models trained by ECT as ECM. In most case, you can refer it as ECT.

Best luck with your journey!

Zhengyang

2019211753 commented 2 months ago

Hi @2019211753 ,

We've updated the environment setup. I suggest setting the number of GPUs as the power of 2. You need a GPU machine with at least 1 A6000 GPU or equivalent. Set up your environment using conda as in README.

Sorry for the confusion. We denote the technique from our blog post as Easy Consistency Tuning (ECT), while the models trained by ECT as ECM. In most case, you can refer it as ECT.

Best luck with your journey!

Zhengyang

Thanks for your reply. It helps me a lot.