Gaurav Parmar
Gaurav Parmar
Hi, Thank you for the interest in the project! Yeah it should be possible to change the method to work with grayscale images! There are two approaches you can take:...
Hi @c7w , Apologies for the delays! The training code for unpaired translation is now uploaded. See the corresponding doc [here](https://github.com/GaParmar/img2img-turbo/blob/main/docs/training_cyclegan_turbo.md). -Gaurav
You can see [this PR](https://github.com/GaParmar/img2img-turbo/pull/30) to reduce the VRAM required with mixed precision training. I trained my unpaired models on A6000, which has 48GB VRAM. -Gaurav
Thank you for documenting this!
Does this issue persist if you try install the `vision_aided_loss` package?
Hi, The input images are scaled to [0,1] and output images are scaled to [-1,1]. See this code block [here](https://github.com/GaParmar/img2img-turbo/blob/main/src/my_utils/training_utils.py#L290).
Hi Jan, Thank you for pointing this out! Feel free to make a PR. Your proposed solution makes a lot of sense, I will add it to the main repo!...
Hi, for our unpaired models we keep the text prompt fixed during both training and inference. Since the model has not seen other text prompts during training, it is hard...
Hi, Thank you for you interest in this project! A couple things I can observe from your results: - What is the image pre-processing used during inference and training time?...
Does the training end early because you set the `max_train_steps=50` in the training command? This issue might be resolved if you increase this to a larger value. -Gaurav