initial impl of multi img/text dataset

comfy_extras/nodes_train.py

@@ -75,7 +75,7 @@ class BiasDiff(torch.nn.Module):
         return self.passive_memory_usage()
 
 
-def load_and_process_images(image_files, input_dir, resize_method="None"):
+def load_and_process_images(image_files, input_dir, resize_method="None", w=None, h=None):
     """Utility function to load and process a list of images.
 
     Args:
@@ -90,7 +90,6 @@ def load_and_process_images(image_files, input_dir, resize_method="None"):
         raise ValueError("No valid images found in input")
 
     output_images = []
-    w, h = None, None
 
     for file in image_files:
         image_path = os.path.join(input_dir, file)
@@ -206,6 +205,88 @@ class LoadImageSetFromFolderNode:
         return (output_tensor,)
 
 
+class LoadImageTextSetFromFolderNode:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "folder": (folder_paths.get_input_subfolders(), {"tooltip": "The folder to load images from."}),
+                "clip": (IO.CLIP, {"tooltip": "The CLIP model used for encoding the text."}),
+            },
+            "optional": {
+                "resize_method": (
+                    ["None", "Stretch", "Crop", "Pad"],
+                    {"default": "None"},
+                ),
+                "width": (
+                    IO.INT,
+                    {
+                        "default": -1,
+                        "min": -1,
+                        "max": 10000,
+                        "step": 1,
+                        "tooltip": "The width to resize the images to. -1 means use the original width.",
+                    },
+                ),
+                "height": (
+                    IO.INT,
+                    {
+                        "default": -1,
+                        "min": -1,
+                        "max": 10000,
+                        "step": 1,
+                        "tooltip": "The height to resize the images to. -1 means use the original height.",
+                    },
+                )
+            },
+        }
+
+    RETURN_TYPES = ("IMAGE", IO.CONDITIONING,)
+    FUNCTION = "load_images"
+    CATEGORY = "loaders"
+    EXPERIMENTAL = True
+    DESCRIPTION = "Loads a batch of images and caption from a directory for training."
+
+    def load_images(self, folder, clip, resize_method, width=None, height=None):
+        if clip is None:
+            raise RuntimeError("ERROR: clip input is invalid: None\n\nIf the clip is from a checkpoint loader node your checkpoint does not contain a valid clip or text encoder model.")
+
+        logging.info(f"Loading images from folder: {folder}")
+
+        sub_input_dir = os.path.join(folder_paths.get_input_directory(), folder)
+        valid_extensions = [".png", ".jpg", ".jpeg", ".webp"]
+        image_files = [
+            f
+            for f in os.listdir(sub_input_dir)
+            if any(f.lower().endswith(ext) for ext in valid_extensions)
+        ]
+        caption_file_path = [
+            os.path.basename(f).replace(os.path.splitext(f)[1], ".txt")
+            for f in image_files
+        ]
+        captions = []
+        for caption_file in caption_file_path:
+            caption_path = os.path.join(sub_input_dir, caption_file)
+            if os.path.exists(caption_path):
+                with open(caption_path, "r", encoding="utf-8") as f:
+                    caption = f.read().strip()
+                    captions.append(caption)
+            else:
+                captions.append("")
+
+        width = width if width != -1 else None
+        height = height if height != -1 else None
+        output_tensor = load_and_process_images(image_files, sub_input_dir, resize_method, width, height)
+
+        logging.info(f"Loaded {len(output_tensor)} images from {sub_input_dir}.")
+
+        conditions = []
+        for text in captions:
+            tokens = clip.tokenize(text)
+            conditions.extend(clip.encode_from_tokens_scheduled(tokens))
+        return (output_tensor, conditions)
+
+
 def draw_loss_graph(loss_map, steps):
     width, height = 500, 300
     img = Image.new("RGB", (width, height), "white")
@@ -381,6 +462,13 @@ class TrainLoraNode:
 
         latents = latents["samples"].to(dtype)
         num_images = latents.shape[0]
+        logging.info(f"Total Images: {num_images}, Total Captions: {len(positive)}")
+        if len(positive) == 1 and num_images > 1:
+            positive = positive * num_images
+        elif len(positive) != num_images:
+            raise ValueError(
+                f"Number of positive conditions ({len(positive)}) does not match number of images ({num_images})."
+            )
 
         with torch.inference_mode(False):
             lora_sd = {}
@@ -474,6 +562,7 @@ class TrainLoraNode:
             # setup models
             for m in find_all_highest_child_module_with_forward(mp.model.diffusion_model):
                 patch(m)
+            mp.model.requires_grad_(False)
             comfy.model_management.load_models_gpu([mp], memory_required=1e20, force_full_load=True)
 
             # Setup sampler and guider like in test script
@@ -486,7 +575,6 @@ class TrainLoraNode:
             )
             guider = comfy_extras.nodes_custom_sampler.Guider_Basic(mp)
             guider.set_conds(positive)  # Set conditioning from input
-            ss = comfy_extras.nodes_custom_sampler.SamplerCustomAdvanced()
 
             # yoland: this currently resize to the first image in the dataset
 
@@ -495,17 +583,17 @@ class TrainLoraNode:
             try:
                 for step in (pbar:=tqdm.trange(steps, desc="Training LoRA", smoothing=0.01, disable=not comfy.utils.PROGRESS_BAR_ENABLED)):
                     # Generate random sigma
-                    sigma = mp.model.model_sampling.percent_to_sigma(
+                    sigmas = [mp.model.model_sampling.percent_to_sigma(
                         torch.rand((1,)).item()
-                    )
-                    sigma = torch.tensor([sigma])
+                    ) for _ in range(min(batch_size, num_images))]
+                    sigmas = torch.tensor(sigmas)
 
                     noise = comfy_extras.nodes_custom_sampler.Noise_RandomNoise(step * 1000 + seed)
 
                     indices = torch.randperm(num_images)[:batch_size]
-                    ss.sample(
-                        noise, guider, train_sampler, sigma, {"samples": latents[indices].clone()}
-                    )
+                    batch_latent = latents[indices].clone()
+                    guider.set_conds([positive[i] for i in indices])  # Set conditioning from input
+                    guider.sample(noise.generate_noise({"samples": batch_latent}), batch_latent, train_sampler, sigmas, seed=noise.seed)
             finally:
                 for m in mp.model.modules():
                     unpatch(m)
@@ -697,6 +785,7 @@ NODE_CLASS_MAPPINGS = {
     "SaveLoRANode": SaveLoRA,
     "LoraModelLoader": LoraModelLoader,
     "LoadImageSetFromFolderNode": LoadImageSetFromFolderNode,
+    "LoadImageTextSetFromFolderNode": LoadImageTextSetFromFolderNode,
     "LossGraphNode": LossGraphNode,
 }
 
@@ -705,5 +794,6 @@ NODE_DISPLAY_NAME_MAPPINGS = {
     "SaveLoRANode": "Save LoRA Weights",
     "LoraModelLoader": "Load LoRA Model",
     "LoadImageSetFromFolderNode": "Load Image Dataset from Folder",
+    "LoadImageTextSetFromFolderNode": "Load Image and Text Dataset from Folder",
     "LossGraphNode": "Plot Loss Graph",
 }