sd3 empty latent, change entrypoint function name

__init__.py

@@ -1,12 +1,15 @@
 from .nodes.sd3_load_checkpoint import *
 from .nodes.sd3_load_clips import *
+from .nodes.sd3_empty_latent import *
 
 NODE_CLASS_MAPPINGS = {
     "SD3LoadCheckpoint": SD3LoadCheckpoint,
     "SD3LoadCLIPs": SD3LoadCLIPs,
+    "SD3EmptyLatent": SD3EmptyLatent,
 }
 
 NODE_DISPLAY_NAME_MAPPINGS = {
     "SD3LoadCheckpoint": "SD3 Load Checkpoint",
     "SD3LoadCLIPs": "SD3 Load CLIPs",
+    "SD3EmptyLatent": "SD3 Empty Latent",
 }

nodes/sd3_empty_latent.py

@@ -0,0 +1,26 @@
+import torch
+import comfy.model_management
+
+class SD3EmptyLatent:
+    def __init__(self):
+        self.device = comfy.model_management.intermediate_device()
+
+    @classmethod
+    def INPUT_TYPES(s):
+        resolution = []
+        # Generate resolutions following this guideline: "Resolution should be around 1 megapixel and width/height must be multiple of 64"
+        for width in range(512, 1921, 64):
+            height = int(1024 * 1024 / width / 64) * 64
+            resolution.append(f"{width}x{height}")
+
+        return {"required": {"resolution": (resolution, {"default": "1024x1024"}),
+                             "batch_size": ("INT", {"default": 1, "min": 1, "max": 16})}}
+    RETURN_TYPES = ("LATENT",)
+    FUNCTION = "main"
+
+    CATEGORY = "latent/sd3"
+
+    def main(self, resolution, batch_size=1):
+        width, height = map(int, resolution.split('x'))
+        latent = torch.zeros([batch_size, 16, height // 8, width // 8], device=self.device)
+        return ({"samples":latent}, )

nodes/sd3_load_checkpoint.py

@@ -9,11 +9,11 @@ class SD3LoadCheckpoint (CheckpointLoaderSimple):
                              "shift": ("FLOAT", {"default": 3.0, "min": 0.0, "max": 100.0, "step":0.01}),
                              }}
     RETURN_TYPES = ("MODEL", "VAE")
-    FUNCTION = "load_checkpoint"
+    FUNCTION = "main"
 
     CATEGORY = "sd3"
 
-    def load_checkpoint(self, ckpt_name, shift):
+    def main(self, ckpt_name, shift):
         ckpt_path = folder_paths.get_full_path("checkpoints", ckpt_name)
         model, _, vae, _ = comfy.sd.load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=False, embedding_directory=folder_paths.get_folder_paths("embeddings"))
         m = model.clone()

nodes/sd3_load_clips.py

@@ -6,7 +6,7 @@ class SD3LoadCLIPs(TripleCLIPLoader):
     def INPUT_TYPES(s):
         return {"required": { "clip_g": (folder_paths.get_filename_list("clip"), ), "clip_l": (folder_paths.get_filename_list("clip"), ), "t5xxl": (folder_paths.get_filename_list("clip"), )}}
 
-    FUNCTION = "load_clips"
+    FUNCTION = "main"
 
-    def load_clips(self, clip_g, clip_l, t5xxl):
+    def main(self, clip_g, clip_l, t5xxl):
         return self.load_clip(clip_g, clip_l, t5xxl)