Confuse about the indices selection in F1 mode

Hello,

Thanks for the great work!!!

I am confuse about the code in demo_gradio_f1.py, about the indices of each inference

        indices = torch.arange(0, sum([1, 16, 2, 1, latent_window_size])).unsqueeze(0)

As the start_latent always be the part of clean_latents, its indices is always 0. However, after multiply inferences, the distance between the start_latent and "clean_latents_4x, clean_latents_2x, clean_latents_1x" should be larger. But there indices is always [0,1,2,......,19].
I think indices should be like [0,10,11,......,28], [0,19,20,.....,37]. The distance between the start_latent and history latents should larger during the inference.

Am i understand correctly?

Thank you.

Start by reading demo_gradio_f1.py around the indices = torch.arange(...) expression and trace how start_latent, clean_latents, and the history latents are assembled during repeated inference. Confirm whether the indices represent positions within each inference window or offsets across windows, and document or correct the behavior so the intended distance is clear.