Vram not flushed after a depth generated, next render produce an out of memory

[MattRM2]

Hi,

I found a bug, when the rendering of a depth map is done, the vram is not flushed. The problem is the next render of the depth map fails because no more memory.

I not using Boost for this exemple (memory error to because my input is too big), manual output resolution 1024x768. First time ok, second time fail because still a lot of vram used that was not flushed.

I used zoeDepth_nk model for my exemple.

I know, I can use cpu (not the same behavior, on cpu the memory is flushed) but I think the flush of the vram can be usefull.

Also, could you tell me the max resolution for GPU and boost to be cut in multiple tile ? One not detailled image in 1920x1080 was computed without problem using boost but with more detailled images, many out of memory error so if you have some information to avoid this error, I take them.

Also I saw that the target resolution is (2353, 3136, 3) but my input is 1277x958. I think I'm missing something, if you know what, tell my what I'm doing wrong. But your stable diffusion depthmap extension is really good and I'm really happy you done something like that, just need to understand what I'm doing wrong and maybe a vram flush button ;)

Thank you

Matt