wisconsinaivision / visii Goto Github PK

Dear authors,

I hope this message finds you well. I'm currently attempting to replicate the Visual Instruction Inversion project you shared on GitHub. However, I couldn't locate any checkpoints related to InstructPix2Pix within the provided codebase.

While going through the 'train.py' file, I noticed that the default checkpoint should be placed in the './logs/' directory. Regrettably, I couldn't find this directory within the project.

I wanted to inquire whether this code includes the checkpoint section. If not, would it be possible for me to request a checkpoint from you to facilitate my replication process?

Thank you sincerely for your time and consideration. I eagerly await your response.

Best regards,
Zou Ling

Hi author, thanks for your team's contribution.

I would like to ask you a question about calculating the metrics during the training process. Specifically, the training process is usually interspersed with a validation step, do you perform the computation of the evaluation metrics during the validation step, which seems to be time consuming. So I'm wondering how you schedule the evaluation during the training process?

Hi I have a quesion about the reproduct.

I trained a model according this commend.
"python train.py --image_folder ./images --subfolder painting1"

And then, I test a model using this commend.
"python test.py --prompt 'a husky'"

However, the results I got are quite different from what the GitHub README suggests.

Which options can I change to achieve the same output?

Thank you!

Hi, what do you mean for clean-instructpix2pix dataset? I see the reference is the original paper, but it doesn't mention clean-instructpix2pix dataset, could you introduce more and give a link?

Hi, thanks for your work. I have some questions as follows:

1. how to perform with controlnet using the following command, as there is no such file named train_controlnet.py

python train_controlnet.py --image_folder ./images --subfolder painting1

2. Could you release a checkpoint that is trained on the combination of the subset of CleanInstructPix2Pix and other data, such that we could reproduce the results of Figure 7, the main result of this paper?

I have followed your tutorial to run "train.py" and "test.py", but no file named "test_concat.py".

Hello,

I'm trying to implement your paper on my own. I'm confused with the "reusing identical noises during inference" part.

During training, since we are randomly sampling timestep t, there could be multiple or no noises sampled for a specific t. For example, with T=1000, during training I could have sampled t=200 twice and never sampled t=100. During inference time, what noises should I use for t=200 and t=100?

Which inference method do you use? If I use deterministic ODE solvers such as DDIM, where no noise is needed during backward process, what should I do?

Do you have an estimated time for code release?

Thank you!

Hey, thanks for the great paper -- its a super interesting idea, do you have an estimated timeline of when you plan to release the code?

I'm very much looking forward to your open source : ) When will you release the code? I am following Visii and doing related research, could you send me a pre-release version for reference? My email is as follows: [email protected]