The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
Faster clip image processing.
comfyanonymous723847f6b3d5
diff --git a/comfy/clip_vision.py b/comfy/clip_vision.pyindex e085186ef..9e2e03d72 100644--- a/comfy/clip_vision.py+++ b/comfy/clip_vision.py@@ -1,5 +1,5 @@-from transformers import CLIPVisionModelWithProjection, CLIPVisionConfig, CLIPImageProcessor, modeling_utils-from .utils import load_torch_file, transformers_convert+from transformers import CLIPVisionModelWithProjection, CLIPVisionConfig, modeling_utils+from .utils import load_torch_file, transformers_convert, common_upscaleimport osimport torchimport contextlib@@ -7,6 +7,18 @@ import contextlibimport comfy.opsimport comfy.model_patcherimport comfy.model_management+import comfy.utils++def clip_preprocess(image, size=224):+ mean = torch.tensor([ 0.48145466,0.4578275,0.40821073], device=image.device, dtype=image.dtype)+ std = torch.tensor([0.26862954,0.26130258,0.27577711], device=image.device, dtype=image.dtype)+ scale = (size / min(image.shape[1], image.shape[2]))+ image = torch.nn.functional.interpolate(image.movedim(-1, 1), size=(round(scale * image.shape[1]), round(scale * image.shape[2])), mode="bicubic", antialias=True)+ h = (image.shape[2] - size)//2+ w = (image.shape[3] - size)//2+ image = image[:,:,h:h+size,w:w+size]+ image = torch.clip((255. * image), 0, 255).round() / 255.0+ return (image - mean.view([3,1,1])) / std.view([3,1,1])class ClipVisionModel():def __init__(self, json_config):@@ -23,25 +35,12 @@ class ClipVisionModel():self.model.to(self.dtype)self.patcher = comfy.model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)- self.processor = CLIPImageProcessor(crop_size=224,- do_center_crop=True,- do_convert_rgb=True,- do_normalize=True,- do_resize=True,- image_mean=[ 0.48145466,0.4578275,0.40821073],- image_std=[0.26862954,0.26130258,0.27577711],- resample=3, #bicubic- size=224)-def load_sd(self, sd):return self.model.load_state_dict(sd, strict=False)def encode_image(self, image):- img = torch.clip((255. * image), 0, 255).round().int()- img = list(map(lambda a: a, img))- inputs = self.processor(images=img, return_tensors="pt")comfy.model_management.load_model_gpu(self.patcher)- pixel_values = inputs['pixel_values'].to(self.load_device)+ pixel_values = clip_preprocess(image.to(self.load_device))if self.dtype != torch.float32:precision_scope = torch.autocast