jaidaken/ComfyUI
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

Support the siglip 2 naflex model as a clip vision model. (#11831)

Browse Source 
Not useful yet.
This commit is contained in:
comfyanonymous
2026-01-12 14:05:54 -08:00
committed by GitHub
parent a3b5d4996a
commit c881a1d689
3 changed files with 91 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

25
comfy/clip_vision.py
View File

@@ -21,6 +21,7 @@ clip_preprocess = comfy.clip_model.clip_preprocess # Prevent some stuff from br
IMAGE_ENCODERS = {
"clip_vision_model": comfy.clip_model.CLIPVisionModelProjection,
"siglip_vision_model": comfy.clip_model.CLIPVisionModelProjection,
"siglip2_vision_model": comfy.clip_model.CLIPVisionModelProjection,
"dinov2": comfy.image_encoders.dino2.Dinov2Model,
}
@@ -32,9 +33,10 @@ class ClipVisionModel():
self.image_size = config.get("image_size", 224)
self.image_mean = config.get("image_mean", [0.48145466, 0.4578275, 0.40821073])
self.image_std = config.get("image_std", [0.26862954, 0.26130258, 0.27577711])
model_type = config.get("model_type", "clip_vision_model")
model_class = IMAGE_ENCODERS.get(model_type)
if model_type == "siglip_vision_model":
self.model_type = config.get("model_type", "clip_vision_model")
self.config = config.copy()
model_class = IMAGE_ENCODERS.get(self.model_type)
if self.model_type == "siglip_vision_model":
self.return_all_hidden_states = True
else:
self.return_all_hidden_states = False
@@ -55,7 +57,10 @@ class ClipVisionModel():
def encode_image(self, image, crop=True):
comfy.model_management.load_model_gpu(self.patcher)
pixel_values = comfy.clip_model.clip_preprocess(image.to(self.load_device), size=self.image_size, mean=self.image_mean, std=self.image_std, crop=crop).float()
if self.model_type == "siglip2_vision_model":
pixel_values = comfy.clip_model.siglip2_preprocess(image.to(self.load_device), size=self.image_size, patch_size=self.config.get("patch_size", 16), num_patches=self.config.get("num_patches", 256), mean=self.image_mean, std=self.image_std, crop=crop).float()
else:
pixel_values = comfy.clip_model.clip_preprocess(image.to(self.load_device), size=self.image_size, mean=self.image_mean, std=self.image_std, crop=crop).float()
out = self.model(pixel_values=pixel_values, intermediate_output='all' if self.return_all_hidden_states else -2)
outputs = Output()
@@ -107,10 +112,14 @@ def load_clipvision_from_sd(sd, prefix="", convert_keys=False):
elif "vision_model.encoder.layers.22.layer_norm1.weight" in sd:
embed_shape = sd["vision_model.embeddings.position_embedding.weight"].shape[0]
if sd["vision_model.encoder.layers.0.layer_norm1.weight"].shape[0] == 1152:
if embed_shape == 729:
json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip_384.json")
elif embed_shape == 1024:
json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip_512.json")
patch_embedding_shape = sd["vision_model.embeddings.patch_embedding.weight"].shape
if len(patch_embedding_shape) == 2:
json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip2_base_naflex.json")
else:
if embed_shape == 729:
json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip_384.json")
elif embed_shape == 1024:
json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip_512.json")
elif embed_shape == 577:
if "multi_modal_projector.linear_1.bias" in sd:
json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_vitl_336_llava.json")