training time cost

[yfyang007]

thanks for this excellent work!
i am reproducing this work and wonder how much iterations finally cost for a img2img task,it was decribed in your paper 20k,but for my machine 20k will cost for a long time. So i wonder if there is a solution to accelerate process.
and below is my 300 step result, it is not promising

[mgui7]

Hi, thanks for your interest in our work! I expect the required training time to vary depending on the complexity of the specific image-to-image task. In general, I’d recommend training for longer to achieve optimal performance, as 300 steps is very likely too few for an image generation network to adapt effectively to a new image-to-image translation task.

If you’re constrained by time, you might consider reducing the batch size (e.g., to 64 or even 32) to speed up training. However, keep in mind that this will have an effect on convergence and final performance.

[yfyang007]

Hi, thanks for your interest in our work! I expect the required training time to vary depending on the complexity of the specific image-to-image task. In general, I’d recommend training for longer to achieve optimal performance, as 300 steps is very likely too few for an image generation network to adapt effectively to a new image-to-image translation task.

If you’re constrained by time, you might consider reducing the batch size (e.g., to 64 or even 32) to speed up training. However, keep in mind that this will have an effect on convergence and final performance.

thanks a lot. i also met the situation that when i tried make some change in x_t computation with extra condition,the training seems to be harder than the original version,i wonder if it is normal or need some extra skiil to adjust?
belowing is training loss and result ,the loss converges in step 12k at original version,but after change it is not,there are original loss curve and change version and some not good result in experiment of change version(left :gt ,right:pred)
thanks again for replying!