xinyun/ComfyUI-KJNodes
1
0
Fork 0
mirror of https://git.datalinker.icu/kijai/ComfyUI-KJNodes.git synced 2025-12-13 23:04:40 +08:00
Code Issues Packages Projects Releases Wiki Activity

Make blur run on GPU

Browse Source 
10x faster.. added toggle too to test, defaults on
This commit is contained in:
kijai 2023-10-17 19:54:14 +03:00
parent 546313b439
commit 73fe2f61d6
1 changed files with 12 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
nodes.py
View File

@ -330,6 +330,7 @@ class GrowMaskWithBlur:
"incremental_expandrate": ("INT", {"default": 0, "min": 0, "max": 100, "step": 1}), "incremental_expandrate": ("INT", {"default": 0, "min": 0, "max": 100, "step": 1}),
"tapered_corners": ("BOOLEAN", {"default": True}), "tapered_corners": ("BOOLEAN", {"default": True}),
"flip_input": ("BOOLEAN", {"default": False}), "flip_input": ("BOOLEAN", {"default": False}),
"use_cuda": ("BOOLEAN", {"default": True}),
"blur_radius": ("INT", { "blur_radius": ("INT", {
"default": 0, "default": 0,
"min": 0, "min": 0,
@ -351,7 +352,7 @@ class GrowMaskWithBlur:
RETURN_NAMES = ("mask", "mask_inverted",) RETURN_NAMES = ("mask", "mask_inverted",)
FUNCTION = "expand_mask" FUNCTION = "expand_mask"
def expand_mask(self, mask, expand, tapered_corners, flip_input, blur_radius, sigma, incremental_expandrate): def expand_mask(self, mask, expand, tapered_corners, flip_input, blur_radius, sigma, incremental_expandrate, use_cuda):
if( flip_input ): if( flip_input ):
mask = 1.0 - mask mask = 1.0 - mask
c = 0 if tapered_corners else 1 c = 0 if tapered_corners else 1
@ -373,8 +374,12 @@ class GrowMaskWithBlur:
expand += abs(incremental_expandrate) # Use abs(growrate) to ensure positive change expand += abs(incremental_expandrate) # Use abs(growrate) to ensure positive change
output = torch.from_numpy(output) output = torch.from_numpy(output)
out.append(output) out.append(output)
blurred = torch.stack(out, dim=0).reshape((-1, 1, mask.shape[-2], mask.shape[-1])).movedim(1, -1).expand(-1, -1, -1, 3) blurred = torch.stack(out, dim=0).reshape((-1, 1, mask.shape[-2], mask.shape[-1])).movedim(1, -1).expand(-1, -1, -1, 3)
if use_cuda:
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
blurred = blurred.to(device) # Move blurred tensor to the GPU
batch_size, height, width, channels = blurred.shape batch_size, height, width, channels = blurred.shape
if blur_radius != 0: if blur_radius != 0:
blurkernel_size = blur_radius * 2 + 1 blurkernel_size = blur_radius * 2 + 1
@ -642,7 +647,7 @@ class VRAM_Debug:
return { return {
"required": { "required": {
"model": ("MODEL",), "model": ("MODEL",),
"empty_cuda_cache": ("BOOLEAN", {"default": False}),
}, },
"optional": { "optional": {
"clip_vision": ("CLIP_VISION", ), "clip_vision": ("CLIP_VISION", ),
@ -653,11 +658,12 @@ class VRAM_Debug:
FUNCTION = "VRAMdebug" FUNCTION = "VRAMdebug"
CATEGORY = "KJNodes" CATEGORY = "KJNodes"
def VRAMdebug(self, model, clip_vision=None): def VRAMdebug(self, model, empty_cuda_cache, clip_vision=None):
freemem_before = comfy.model_management.get_free_memory() freemem_before = comfy.model_management.get_free_memory()
print(freemem_before) print(freemem_before)
torch.cuda.empty_cache() if empty_cuda_cache:
torch.cuda.ipc_collect() torch.cuda.empty_cache()
torch.cuda.ipc_collect()
if clip_vision is not None: if clip_vision is not None:
print("unloading clip_vision_clone") print("unloading clip_vision_clone")
comfy.model_management.unload_model_clones(clip_vision.patcher) comfy.model_management.unload_model_clones(clip_vision.patcher)