Initial commit: save/load latent nodes with absolute path support

Preserves full LATENT dict including tensors, non-tensor metadata, and original device placement.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

__init__.py

@@ -0,0 +1,13 @@
+from .nodes import SaveLatentAbsolute, LoadLatentAbsolute
+
+NODE_CLASS_MAPPINGS = {
+    "SaveLatentAbsolute": SaveLatentAbsolute,
+    "LoadLatentAbsolute": LoadLatentAbsolute,
+}
+
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "SaveLatentAbsolute": "Save Latent (Absolute Path)",
+    "LoadLatentAbsolute": "Load Latent (Absolute Path)",
+}
+
+__all__ = ["NODE_CLASS_MAPPINGS", "NODE_DISPLAY_NAME_MAPPINGS"]

nodes.py

@@ -0,0 +1,77 @@
+import os
+import json
+import torch
+import safetensors.torch
+
+
+class SaveLatentAbsolute:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "samples": ("LATENT",),
+                "path": ("STRING", {"default": "/path/to/latent.latent"}),
+            }
+        }
+
+    RETURN_TYPES = ("LATENT",)
+    FUNCTION = "save"
+    CATEGORY = "latent"
+    OUTPUT_NODE = True
+
+    def save(self, samples, path):
+        path = os.path.expanduser(path)
+        os.makedirs(os.path.dirname(path), exist_ok=True)
+
+        tensors = {}
+        non_tensors = {}
+        devices = {}
+        for key, value in samples.items():
+            if isinstance(value, torch.Tensor):
+                devices[key] = str(value.device)
+                tensors[key] = value.contiguous()
+            else:
+                non_tensors[key] = value
+
+        metadata = {"devices": json.dumps(devices)}
+        if non_tensors:
+            metadata["non_tensor_data"] = json.dumps(non_tensors)
+
+        safetensors.torch.save_file(tensors, path, metadata=metadata)
+        return (samples,)
+
+
+class LoadLatentAbsolute:
+    @classmethod
+    def INPUT_TYPES(cls):
+        return {
+            "required": {
+                "path": ("STRING", {"default": "/path/to/latent.latent"}),
+            }
+        }
+
+    RETURN_TYPES = ("LATENT",)
+    FUNCTION = "load"
+    CATEGORY = "latent"
+
+    def load(self, path):
+        path = os.path.expanduser(path)
+
+        samples = safetensors.torch.load_file(path, device="cpu")
+
+        with safetensors.safe_open(path, framework="pt") as f:
+            meta = f.metadata()
+
+        # Restore original devices
+        if meta and "devices" in meta:
+            devices = json.loads(meta["devices"])
+            for key, device in devices.items():
+                if key in samples:
+                    samples[key] = samples[key].to(device)
+
+        # Restore non-tensor data
+        if meta and "non_tensor_data" in meta:
+            non_tensors = json.loads(meta["non_tensor_data"])
+            samples.update(non_tensors)
+
+        return (samples,)