Add unfinished ImageOnlyCheckpointSave node to save a SVD checkpoint.

This node is unfinished, SVD checkpoints saved with this node will
work with ComfyUI but not with anything else.

comfy/clip_vision.py

@@ -1,4 +1,4 @@
-from .utils import load_torch_file, transformers_convert, common_upscale
+from .utils import load_torch_file, transformers_convert, common_upscale, state_dict_prefix_replace
 import os
 import torch
 import contextlib
@@ -41,9 +41,13 @@ class ClipVisionModel():
         self.model.eval()
 
         self.patcher = comfy.model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
+
     def load_sd(self, sd):
         return self.model.load_state_dict(sd, strict=False)
 
+    def get_sd(self):
+        return self.model.state_dict()
+
     def encode_image(self, image):
         comfy.model_management.load_model_gpu(self.patcher)
         pixel_values = clip_preprocess(image.to(self.load_device)).float()
@@ -76,6 +80,9 @@ def convert_to_transformers(sd, prefix):
             sd['visual_projection.weight'] = sd.pop("{}proj".format(prefix)).transpose(0, 1)
 
         sd = transformers_convert(sd, prefix, "vision_model.", 48)
+    else:
+        replace_prefix = {prefix: ""}
+        sd = state_dict_prefix_replace(sd, replace_prefix)
     return sd
 
 def load_clipvision_from_sd(sd, prefix="", convert_keys=False):

comfy/model_base.py

@@ -179,19 +179,28 @@ class BaseModel(torch.nn.Module):
     def process_latent_out(self, latent):
         return self.latent_format.process_out(latent)
 
-    def state_dict_for_saving(self, clip_state_dict, vae_state_dict):
-        clip_state_dict = self.model_config.process_clip_state_dict_for_saving(clip_state_dict)
+    def state_dict_for_saving(self, clip_state_dict=None, vae_state_dict=None, clip_vision_state_dict=None):
+        extra_sds = []
+        if clip_state_dict is not None:
+            extra_sds.append(self.model_config.process_clip_state_dict_for_saving(clip_state_dict))
+        if vae_state_dict is not None:
+            extra_sds.append(self.model_config.process_vae_state_dict_for_saving(vae_state_dict))
+        if clip_vision_state_dict is not None:
+            extra_sds.append(self.model_config.process_clip_vision_state_dict_for_saving(clip_vision_state_dict))
+
         unet_state_dict = self.diffusion_model.state_dict()
         unet_state_dict = self.model_config.process_unet_state_dict_for_saving(unet_state_dict)
-        vae_state_dict = self.model_config.process_vae_state_dict_for_saving(vae_state_dict)
+
         if self.get_dtype() == torch.float16:
-            clip_state_dict = utils.convert_sd_to(clip_state_dict, torch.float16)
-            vae_state_dict = utils.convert_sd_to(vae_state_dict, torch.float16)
+            extra_sds = map(lambda sd: utils.convert_sd_to(sd, torch.float16), extra_sds)
 
         if self.model_type == ModelType.V_PREDICTION:
             unet_state_dict["v_pred"] = torch.tensor([])
 
-        return {**unet_state_dict, **vae_state_dict, **clip_state_dict}
+        for sd in extra_sds:
+            unet_state_dict.update(sd)
+
+        return unet_state_dict
 
     def set_inpaint(self):
         self.inpaint_model = True

comfy/sd.py

@@ -534,7 +534,14 @@ def load_unet(unet_path):
         raise RuntimeError("ERROR: Could not detect model type of: {}".format(unet_path))
     return model
 
-def save_checkpoint(output_path, model, clip, vae, metadata=None):
-    model_management.load_models_gpu([model, clip.load_model()])
-    sd = model.model.state_dict_for_saving(clip.get_sd(), vae.get_sd())
+def save_checkpoint(output_path, model, clip=None, vae=None, clip_vision=None, metadata=None):
+    clip_sd = None
+    load_models = [model]
+    if clip is not None:
+        load_models.append(clip.load_model())
+        clip_sd = clip.get_sd()
+
+    model_management.load_models_gpu(load_models)
+    clip_vision_sd = clip_vision.get_sd() if clip_vision is not None else None
+    sd = model.model.state_dict_for_saving(clip_sd, vae.get_sd(), clip_vision_sd)
     comfy.utils.save_torch_file(sd, output_path, metadata=metadata)

comfy/supported_models_base.py

@@ -65,6 +65,12 @@ class BASE:
         replace_prefix = {"": "cond_stage_model."}
         return utils.state_dict_prefix_replace(state_dict, replace_prefix)
 
+    def process_clip_vision_state_dict_for_saving(self, state_dict):
+        replace_prefix = {}
+        if self.clip_vision_prefix is not None:
+            replace_prefix[""] = self.clip_vision_prefix
+        return utils.state_dict_prefix_replace(state_dict, replace_prefix)
+
     def process_unet_state_dict_for_saving(self, state_dict):
         replace_prefix = {"": "model.diffusion_model."}
         return utils.state_dict_prefix_replace(state_dict, replace_prefix)