diff --git a/flux_train_network.py b/flux_train_network.py
index 34b2be80..67eacefc 100644
--- a/flux_train_network.py
+++ b/flux_train_network.py
@@ -327,14 +327,14 @@ class FluxNetworkTrainer(train_network.NetworkTrainer):
         bsz = latents.shape[0]
 
         # Get CDC parameters if enabled
-        gamma_b_dataset = self.gamma_b_dataset if (self.gamma_b_dataset is not None and "image_keys" in batch) else None
-        image_keys = batch.get("image_keys") if gamma_b_dataset is not None else None
+        gamma_b_dataset = self.gamma_b_dataset if (self.gamma_b_dataset is not None and "latents_npz" in batch) else None
+        latents_npz_paths = batch.get("latents_npz") if gamma_b_dataset is not None else None
 
         # Get noisy model input and timesteps
         # If CDC is enabled, this will transform the noise with geometry-aware covariance
         noisy_model_input, timesteps, sigmas = flux_train_utils.get_noisy_model_input_and_timesteps(
             args, noise_scheduler, latents, noise, accelerator.device, weight_dtype,
-            gamma_b_dataset=gamma_b_dataset, image_keys=image_keys
+            gamma_b_dataset=gamma_b_dataset, latents_npz_paths=latents_npz_paths
         )
 
         # pack latents and get img_ids
diff --git a/library/cdc_fm.py b/library/cdc_fm.py
index 10b00864..84a8a34a 100644
--- a/library/cdc_fm.py
+++ b/library/cdc_fm.py
@@ -7,12 +7,6 @@ from safetensors.torch import save_file
 from typing import List, Dict, Optional, Union, Tuple
 from dataclasses import dataclass
 
-try:
-    import faiss  # type: ignore
-    FAISS_AVAILABLE = True
-except ImportError:
-    FAISS_AVAILABLE = False
-
 logger = logging.getLogger(__name__)
 
 
@@ -24,6 +18,7 @@ class LatentSample:
     latent: np.ndarray  # (d,) flattened latent vector
     global_idx: int  # Global index in dataset
     shape: Tuple[int, ...]  # Original shape before flattening (C, H, W)
+    latents_npz_path: str  # Path to the latent cache file
     metadata: Optional[Dict] = None  # Any extra info (prompt, filename, etc.)
 
 
@@ -49,7 +44,7 @@ class CarreDuChampComputer:
 
     def compute_knn_graph(self, latents_np: np.ndarray) -> Tuple[np.ndarray, np.ndarray]:
         """
-        Build k-NN graph using FAISS
+        Build k-NN graph using pure PyTorch
 
         Args:
             latents_np: (N, d) numpy array of same-dimensional latents
@@ -63,19 +58,48 @@ class CarreDuChampComputer:
         # Clamp k to available neighbors (can't have more neighbors than samples)
         k_actual = min(self.k, N - 1)
 
-        # Ensure float32
-        if latents_np.dtype != np.float32:
-            latents_np = latents_np.astype(np.float32)
+        # Convert to torch tensor
+        latents_tensor = torch.from_numpy(latents_np).to(self.device)
 
-        # Build FAISS index
-        index = faiss.IndexFlatL2(d)
+        # Compute pairwise L2 distances efficiently
+        # ||a - b||^2 = ||a||^2 + ||b||^2 - 2<a, b>
+        # This is more memory efficient than computing all pairwise differences
+        # For large batches, we'll chunk the computation
+        chunk_size = 1000  # Process 1000 queries at a time to manage memory
 
-        if torch.cuda.is_available():
-            res = faiss.StandardGpuResources()
-            index = faiss.index_cpu_to_gpu(res, 0, index)
+        if N <= chunk_size:
+            # Small batch: compute all at once
+            distances_sq = torch.cdist(latents_tensor, latents_tensor, p=2) ** 2
+            distances_k_sq, indices_k = torch.topk(
+                distances_sq, k=k_actual + 1, dim=1, largest=False
+            )
+            distances = torch.sqrt(distances_k_sq).cpu().numpy()
+            indices = indices_k.cpu().numpy()
+        else:
+            # Large batch: chunk to avoid OOM
+            distances_list = []
+            indices_list = []
 
-        index.add(latents_np)  # type: ignore
-        distances, indices = index.search(latents_np, k_actual + 1)  # type: ignore
+            for i in range(0, N, chunk_size):
+                end_i = min(i + chunk_size, N)
+                chunk = latents_tensor[i:end_i]
+
+                # Compute distances for this chunk
+                distances_sq = torch.cdist(chunk, latents_tensor, p=2) ** 2
+                distances_k_sq, indices_k = torch.topk(
+                    distances_sq, k=k_actual + 1, dim=1, largest=False
+                )
+
+                distances_list.append(torch.sqrt(distances_k_sq).cpu().numpy())
+                indices_list.append(indices_k.cpu().numpy())
+
+                # Free memory
+                del distances_sq, distances_k_sq, indices_k
+                if torch.cuda.is_available():
+                    torch.cuda.empty_cache()
+
+            distances = np.concatenate(distances_list, axis=0)
+            indices = np.concatenate(indices_list, axis=0)
 
         return distances, indices
 
@@ -312,15 +336,17 @@ class LatentBatcher:
         self,
         latent: Union[np.ndarray, torch.Tensor],
         global_idx: int,
+        latents_npz_path: str,
         shape: Optional[Tuple[int, ...]] = None,
         metadata: Optional[Dict] = None
     ):
         """
         Add a latent vector with automatic shape tracking
-        
+
         Args:
             latent: Latent vector (any shape, will be flattened)
             global_idx: Global index in dataset
+            latents_npz_path: Path to the latent cache file (e.g., "image_0512x0768_flux.npz")
             shape: Original shape (if None, uses latent.shape)
             metadata: Optional metadata dict
         """
@@ -337,6 +363,7 @@ class LatentBatcher:
             latent=latent_flat,
             global_idx=global_idx,
             shape=original_shape,
+            latents_npz_path=latents_npz_path,
             metadata=metadata
         )
 
@@ -443,15 +470,9 @@ class CDCPreprocessor:
         size_tolerance: float = 0.0,
         debug: bool = False,
         adaptive_k: bool = False,
-        min_bucket_size: int = 16
+        min_bucket_size: int = 16,
+        dataset_dirs: Optional[List[str]] = None
     ):
-        if not FAISS_AVAILABLE:
-            raise ImportError(
-                "FAISS is required for CDC-FM but not installed. "
-                "Install with: pip install faiss-cpu (CPU) or faiss-gpu (GPU). "
-                "CDC-FM will be disabled."
-            )
-
         self.computer = CarreDuChampComputer(
             k_neighbors=k_neighbors,
             k_bandwidth=k_bandwidth,
@@ -463,37 +484,88 @@ class CDCPreprocessor:
         self.debug = debug
         self.adaptive_k = adaptive_k
         self.min_bucket_size = min_bucket_size
+        self.dataset_dirs = dataset_dirs or []
+        self.config_hash = self._compute_config_hash()
+
+    def _compute_config_hash(self) -> str:
+        """
+        Compute a short hash of CDC configuration for filename uniqueness.
+
+        Hash includes:
+        - Sorted dataset/subset directory paths
+        - CDC parameters (k_neighbors, d_cdc, gamma)
+
+        This ensures CDC files are invalidated when:
+        - Dataset composition changes (different dirs)
+        - CDC parameters change
+
+        Returns:
+            8-character hex hash
+        """
+        import hashlib
+
+        # Sort dataset dirs for consistent hashing
+        dirs_str = "|".join(sorted(self.dataset_dirs))
+
+        # Include CDC parameters
+        config_str = f"{dirs_str}|k={self.computer.k}|d={self.computer.d_cdc}|gamma={self.computer.gamma}"
+
+        # Create short hash (8 chars is enough for uniqueness in this context)
+        hash_obj = hashlib.sha256(config_str.encode())
+        return hash_obj.hexdigest()[:8]
 
     def add_latent(
         self,
         latent: Union[np.ndarray, torch.Tensor],
         global_idx: int,
+        latents_npz_path: str,
         shape: Optional[Tuple[int, ...]] = None,
         metadata: Optional[Dict] = None
     ):
         """
         Add a single latent to the preprocessing queue
-        
+
         Args:
             latent: Latent vector (will be flattened)
             global_idx: Global dataset index
+            latents_npz_path: Path to the latent cache file
             shape: Original shape (C, H, W)
             metadata: Optional metadata
         """
-        self.batcher.add_latent(latent, global_idx, shape, metadata)
+        self.batcher.add_latent(latent, global_idx, latents_npz_path, shape, metadata)
 
-    def compute_all(self, save_path: Union[str, Path]) -> Path:
+    @staticmethod
+    def get_cdc_npz_path(latents_npz_path: str, config_hash: Optional[str] = None) -> str:
         """
-        Compute Γ_b for all added latents and save to safetensors
-        
+        Get CDC cache path from latents cache path
+
+        Includes optional config_hash to ensure CDC files are unique to dataset/subset
+        configuration and CDC parameters. This prevents using stale CDC files when
+        the dataset composition or CDC settings change.
+
         Args:
-            save_path: Path to save the results
-            
+            latents_npz_path: Path to latent cache (e.g., "image_0512x0768_flux.npz")
+            config_hash: Optional 8-char hash of (dataset_dirs + CDC params)
+                        If None, returns path without hash (for backward compatibility)
+
         Returns:
-            Path to saved file
+            CDC cache path:
+            - With hash: "image_0512x0768_flux_cdc_a1b2c3d4.npz"
+            - Without: "image_0512x0768_flux_cdc.npz"
+        """
+        path = Path(latents_npz_path)
+        if config_hash:
+            return str(path.with_stem(f"{path.stem}_cdc_{config_hash}"))
+        else:
+            return str(path.with_stem(f"{path.stem}_cdc"))
+
+    def compute_all(self) -> int:
+        """
+        Compute Γ_b for all added latents and save individual CDC files next to each latent cache
+
+        Returns:
+            Number of CDC files saved
         """
-        save_path = Path(save_path)
-        save_path.parent.mkdir(parents=True, exist_ok=True)
 
         # Get batches by exact size (no resizing)
         batches = self.batcher.get_batches()
@@ -603,90 +675,86 @@ class CDCPreprocessor:
             # Merge into overall results
             all_results.update(batch_results)
 
-        # Save to safetensors
+        # Save individual CDC files next to each latent cache
         if self.debug:
             print(f"\n{'='*60}")
-            print("Saving results...")
+            print("Saving individual CDC files...")
             print(f"{'='*60}")
 
-        tensors_dict = {
-            'metadata/num_samples': torch.tensor([len(all_results)]),
-            'metadata/k_neighbors': torch.tensor([self.computer.k]),
-            'metadata/d_cdc': torch.tensor([self.computer.d_cdc]),
-            'metadata/gamma': torch.tensor([self.computer.gamma]),
-        }
+        files_saved = 0
+        total_size = 0
 
-        # Add shape information and CDC results for each sample
-        # Use image_key as the identifier
-        for sample in self.batcher.samples:
-            image_key = sample.metadata['image_key']
-            tensors_dict[f'shapes/{image_key}'] = torch.tensor(sample.shape)
+        save_iter = tqdm(self.batcher.samples, desc="Saving CDC files", disable=self.debug) if not self.debug else self.batcher.samples
+
+        for sample in save_iter:
+            # Get CDC cache path with config hash
+            cdc_path = self.get_cdc_npz_path(sample.latents_npz_path, self.config_hash)
 
             # Get CDC results for this sample
             if sample.global_idx in all_results:
                 eigvecs, eigvals = all_results[sample.global_idx]
 
-                # Convert numpy arrays to torch tensors
-                if isinstance(eigvecs, np.ndarray):
-                    eigvecs = torch.from_numpy(eigvecs)
-                if isinstance(eigvals, np.ndarray):
-                    eigvals = torch.from_numpy(eigvals)
+                # Convert to numpy if needed
+                if isinstance(eigvecs, torch.Tensor):
+                    eigvecs = eigvecs.numpy()
+                if isinstance(eigvals, torch.Tensor):
+                    eigvals = eigvals.numpy()
 
-                tensors_dict[f'eigenvectors/{image_key}'] = eigvecs
-                tensors_dict[f'eigenvalues/{image_key}'] = eigvals
+                # Save metadata and CDC results
+                np.savez(
+                    cdc_path,
+                    eigenvectors=eigvecs,
+                    eigenvalues=eigvals,
+                    shape=np.array(sample.shape),
+                    k_neighbors=self.computer.k,
+                    d_cdc=self.computer.d_cdc,
+                    gamma=self.computer.gamma
+                )
 
-        save_file(tensors_dict, save_path)
+                files_saved += 1
+                total_size += Path(cdc_path).stat().st_size
 
-        file_size_gb = save_path.stat().st_size / 1024 / 1024 / 1024
-        logger.info(f"Saved to {save_path}")
-        logger.info(f"File size: {file_size_gb:.2f} GB")
+                logger.debug(f"Saved CDC file: {cdc_path}")
 
-        return save_path
+        total_size_mb = total_size / 1024 / 1024
+        logger.info(f"Saved {files_saved} CDC files, total size: {total_size_mb:.2f} MB")
+
+        return files_saved
 
 
 class GammaBDataset:
     """
     Efficient loader for Γ_b matrices during training
-    Handles variable-size latents
+    Loads from individual CDC cache files next to latent caches
     """
 
-    def __init__(self, gamma_b_path: Union[str, Path], device: str = 'cuda'):
+    def __init__(self, device: str = 'cuda', config_hash: Optional[str] = None):
+        """
+        Initialize CDC dataset loader
+
+        Args:
+            device: Device for loading tensors
+            config_hash: Optional config hash to use for CDC file lookup.
+                        If None, uses default naming without hash.
+        """
         self.device = torch.device(device if torch.cuda.is_available() else 'cpu')
-        self.gamma_b_path = Path(gamma_b_path)
-
-        # Load metadata
-        logger.info(f"Loading Γ_b from {gamma_b_path}...")
-        from safetensors import safe_open
-
-        with safe_open(str(self.gamma_b_path), framework="pt", device="cpu") as f:
-            self.num_samples = int(f.get_tensor('metadata/num_samples').item())
-            self.d_cdc = int(f.get_tensor('metadata/d_cdc').item())
-
-            # Cache all shapes in memory to avoid repeated I/O during training
-            # Loading once at init is much faster than opening the file every training step
-            self.shapes_cache = {}
-            # Get all shape keys (they're stored as shapes/{image_key})
-            all_keys = f.keys()
-            shape_keys = [k for k in all_keys if k.startswith('shapes/')]
-            for shape_key in shape_keys:
-                image_key = shape_key.replace('shapes/', '')
-                shape_tensor = f.get_tensor(shape_key)
-                self.shapes_cache[image_key] = tuple(shape_tensor.numpy().tolist())
-
-        logger.info(f"Loaded CDC data for {self.num_samples} samples (d_cdc={self.d_cdc})")
-        logger.info(f"Cached {len(self.shapes_cache)} shapes in memory")
+        self.config_hash = config_hash
+        if config_hash:
+            logger.info(f"CDC loader initialized (hash: {config_hash})")
+        else:
+            logger.info("CDC loader initialized (no hash, backward compatibility mode)")
 
     @torch.no_grad()
     def get_gamma_b_sqrt(
         self,
-        image_keys: Union[List[str], List],
+        latents_npz_paths: List[str],
         device: Optional[str] = None
     ) -> Tuple[torch.Tensor, torch.Tensor]:
         """
-        Get Γ_b^(1/2) components for a batch of image_keys
+        Get Γ_b^(1/2) components for a batch of latents
 
         Args:
-            image_keys: List of image_key strings
+            latents_npz_paths: List of latent cache paths (e.g., ["image_0512x0768_flux.npz", ...])
             device: Device to load to (defaults to self.device)
 
         Returns:
@@ -696,19 +764,26 @@ class GammaBDataset:
         if device is None:
             device = self.device
 
-        # Load from safetensors
-        from safetensors import safe_open
-
         eigenvectors_list = []
         eigenvalues_list = []
 
-        with safe_open(str(self.gamma_b_path), framework="pt", device=str(device)) as f:
-            for image_key in image_keys:
-                eigvecs = f.get_tensor(f'eigenvectors/{image_key}').float()
-                eigvals = f.get_tensor(f'eigenvalues/{image_key}').float()
+        for latents_npz_path in latents_npz_paths:
+            # Get CDC cache path with config hash
+            cdc_path = CDCPreprocessor.get_cdc_npz_path(latents_npz_path, self.config_hash)
 
-                eigenvectors_list.append(eigvecs)
-                eigenvalues_list.append(eigvals)
+            # Load CDC data
+            if not Path(cdc_path).exists():
+                raise FileNotFoundError(
+                    f"CDC cache file not found: {cdc_path}. "
+                    f"Make sure to run CDC preprocessing before training."
+                )
+
+            data = np.load(cdc_path)
+            eigvecs = torch.from_numpy(data['eigenvectors']).to(device).float()
+            eigvals = torch.from_numpy(data['eigenvalues']).to(device).float()
+
+            eigenvectors_list.append(eigvecs)
+            eigenvalues_list.append(eigvals)
 
         # Stack - all should have same d_cdc and d within a batch (enforced by bucketing)
         # Check if all eigenvectors have the same dimension
@@ -718,7 +793,7 @@ class GammaBDataset:
             # but can occur if batch contains mixed sizes
             raise RuntimeError(
                 f"CDC eigenvector dimension mismatch in batch: {set(dims)}. "
-                f"Image keys: {image_keys}. "
+                f"Latent paths: {latents_npz_paths}. "
                 f"This means the training batch contains images of different sizes, "
                 f"which violates CDC's requirement for uniform latent dimensions per batch. "
                 f"Check that your dataloader buckets are configured correctly."
@@ -729,10 +804,6 @@ class GammaBDataset:
 
         return eigenvectors, eigenvalues
 
-    def get_shape(self, image_key: str) -> Tuple[int, ...]:
-        """Get the original shape for a sample (cached in memory)"""
-        return self.shapes_cache[image_key]
-
     def compute_sigma_t_x(
         self,
         eigenvectors: torch.Tensor,
diff --git a/library/flux_train_utils.py b/library/flux_train_utils.py
index 6286ba5b..e503a60e 100644
--- a/library/flux_train_utils.py
+++ b/library/flux_train_utils.py
@@ -476,7 +476,7 @@ def apply_cdc_noise_transformation(
     timesteps: torch.Tensor,
     num_timesteps: int,
     gamma_b_dataset,
-    image_keys,
+    latents_npz_paths,
     device
 ) -> torch.Tensor:
     """
@@ -487,7 +487,7 @@ def apply_cdc_noise_transformation(
         timesteps: (B,) timesteps for this batch
         num_timesteps: Total number of timesteps in scheduler
         gamma_b_dataset: GammaBDataset with cached CDC matrices
-        image_keys: List of image_key strings for this batch
+        latents_npz_paths: List of latent cache paths for this batch
         device: Device to load CDC matrices to
 
     Returns:
@@ -517,62 +517,24 @@ def apply_cdc_noise_transformation(
     t_normalized = timesteps.to(device) / num_timesteps
 
     B, C, H, W = noise.shape
-    current_shape = (C, H, W)
 
-    # Fast path: Check if all samples have matching shapes (common case)
-    # This avoids per-sample processing when bucketing is consistent
-    cached_shapes = [gamma_b_dataset.get_shape(image_key) for image_key in image_keys]
-
-    all_match = all(s == current_shape for s in cached_shapes)
-
-    if all_match:
-        # Batch processing: All shapes match, process entire batch at once
-        eigvecs, eigvals = gamma_b_dataset.get_gamma_b_sqrt(image_keys, device=device)
-        noise_flat = noise.reshape(B, -1)
-        noise_cdc_flat = gamma_b_dataset.compute_sigma_t_x(eigvecs, eigvals, noise_flat, t_normalized)
-        return noise_cdc_flat.reshape(B, C, H, W)
-    else:
-        # Slow path: Some shapes mismatch, process individually
-        noise_transformed = []
-
-        for i in range(B):
-            image_key = image_keys[i]
-            cached_shape = cached_shapes[i]
-
-            if cached_shape != current_shape:
-                # Shape mismatch - use standard Gaussian noise for this sample
-                # Only warn once per sample to avoid log spam
-                if image_key not in _cdc_warned_samples:
-                    logger.warning(
-                        f"CDC shape mismatch for sample {image_key}: "
-                        f"cached {cached_shape} vs current {current_shape}. "
-                        f"Using Gaussian noise (no CDC)."
-                    )
-                    _cdc_warned_samples.add(image_key)
-                noise_transformed.append(noise[i].clone())
-            else:
-                # Shapes match - apply CDC transformation
-                eigvecs, eigvals = gamma_b_dataset.get_gamma_b_sqrt([image_key], device=device)
-
-                noise_flat = noise[i].reshape(1, -1)
-                t_single = t_normalized[i:i+1] if t_normalized.dim() > 0 else t_normalized
-
-                noise_cdc_flat = gamma_b_dataset.compute_sigma_t_x(eigvecs, eigvals, noise_flat, t_single)
-                noise_transformed.append(noise_cdc_flat.reshape(C, H, W))
-
-        return torch.stack(noise_transformed, dim=0)
+    # Batch processing: Get CDC data for all samples at once
+    eigvecs, eigvals = gamma_b_dataset.get_gamma_b_sqrt(latents_npz_paths, device=device)
+    noise_flat = noise.reshape(B, -1)
+    noise_cdc_flat = gamma_b_dataset.compute_sigma_t_x(eigvecs, eigvals, noise_flat, t_normalized)
+    return noise_cdc_flat.reshape(B, C, H, W)
 
 
 def get_noisy_model_input_and_timesteps(
     args, noise_scheduler, latents: torch.Tensor, noise: torch.Tensor, device, dtype,
-    gamma_b_dataset=None, image_keys=None
+    gamma_b_dataset=None, latents_npz_paths=None
 ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
     """
     Get noisy model input and timesteps for training.
 
     Args:
         gamma_b_dataset: Optional CDC-FM gamma_b dataset for geometry-aware noise
-        image_keys: Optional list of image_key strings for CDC-FM (required if gamma_b_dataset provided)
+        latents_npz_paths: Optional list of latent cache file paths for CDC-FM (required if gamma_b_dataset provided)
     """
     bsz, _, h, w = latents.shape
     assert bsz > 0, "Batch size not large enough"
@@ -618,13 +580,13 @@ def get_noisy_model_input_and_timesteps(
     sigmas = sigmas.view(-1, 1, 1, 1)
 
     # Apply CDC-FM geometry-aware noise transformation if enabled
-    if gamma_b_dataset is not None and image_keys is not None:
+    if gamma_b_dataset is not None and latents_npz_paths is not None:
         noise = apply_cdc_noise_transformation(
             noise=noise,
             timesteps=timesteps,
             num_timesteps=num_timesteps,
             gamma_b_dataset=gamma_b_dataset,
-            image_keys=image_keys,
+            latents_npz_paths=latents_npz_paths,
             device=device
         )
 
diff --git a/library/train_util.py b/library/train_util.py
index 9934a52e..a06fc4ef 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -40,6 +40,8 @@ from torch.optim import Optimizer
 from torchvision import transforms
 from transformers import CLIPTokenizer, CLIPTextModel, CLIPTextModelWithProjection
 import transformers
+
+from library.cdc_fm import CDCPreprocessor
 from diffusers.optimization import (
     SchedulerType as DiffusersSchedulerType,
     TYPE_TO_SCHEDULER_FUNCTION as DIFFUSERS_TYPE_TO_SCHEDULER_FUNCTION,
@@ -1570,13 +1572,15 @@ class BaseDataset(torch.utils.data.Dataset):
         text_encoder_outputs_list = []
         custom_attributes = []
         image_keys = []  # CDC-FM: track image keys for CDC lookup
+        latents_npz_paths = []  # CDC-FM: track latents_npz paths for CDC lookup
 
         for image_key in bucket[image_index : image_index + bucket_batch_size]:
             image_info = self.image_data[image_key]
             subset = self.image_to_subset[image_key]
 
-            # CDC-FM: Store image_key for CDC lookup
+            # CDC-FM: Store image_key and latents_npz path for CDC lookup
             image_keys.append(image_key)
+            latents_npz_paths.append(image_info.latents_npz)
 
             custom_attributes.append(subset.custom_attributes)
 
@@ -1823,8 +1827,8 @@ class BaseDataset(torch.utils.data.Dataset):
 
         example["network_multipliers"] = torch.FloatTensor([self.network_multiplier] * len(captions))
 
-        # CDC-FM: Add image keys to batch for CDC lookup
-        example["image_keys"] = image_keys
+        # CDC-FM: Add latents_npz paths to batch for CDC lookup
+        example["latents_npz"] = latents_npz_paths
 
         if self.debug_dataset:
             example["image_keys"] = bucket[image_index : image_index + self.batch_size]
@@ -2709,12 +2713,15 @@ class DatasetGroup(torch.utils.data.ConcatDataset):
         debug: bool = False,
         adaptive_k: bool = False,
         min_bucket_size: int = 16,
-    ) -> str:
+    ) -> Optional[str]:
         """
         Cache CDC Γ_b matrices for all latents in the dataset
 
+        CDC files are saved as individual .npz files next to each latent cache file.
+        For example: image_0512x0768_flux.npz → image_0512x0768_flux_cdc.npz
+
         Args:
-            cdc_output_path: Path to save cdc_gamma_b.safetensors
+            cdc_output_path: Deprecated (CDC uses per-file caching now)
             k_neighbors: k-NN neighbors
             k_bandwidth: Bandwidth estimation neighbors
             d_cdc: CDC subspace dimension
@@ -2723,45 +2730,54 @@ class DatasetGroup(torch.utils.data.ConcatDataset):
             accelerator: For multi-GPU support
 
         Returns:
-            Path to cached CDC file
+            "per_file" to indicate per-file caching is used, or None on error
         """
         from pathlib import Path
 
-        cdc_path = Path(cdc_output_path)
+        # Collect dataset/subset directories for config hash
+        dataset_dirs = []
+        for dataset in self.datasets:
+            # Get the directory containing the images
+            if hasattr(dataset, 'image_dir'):
+                dataset_dirs.append(str(dataset.image_dir))
+            # Fallback: use first image's parent directory
+            elif dataset.image_data:
+                first_image = next(iter(dataset.image_data.values()))
+                dataset_dirs.append(str(Path(first_image.absolute_path).parent))
 
-        # Check if valid cache exists
-        if cdc_path.exists() and not force_recache:
-            if self._is_cdc_cache_valid(cdc_path, k_neighbors, d_cdc, gamma):
-                logger.info(f"Valid CDC cache found at {cdc_path}, skipping preprocessing")
-                return str(cdc_path)
+        # Create preprocessor to get config hash
+        preprocessor = CDCPreprocessor(
+            k_neighbors=k_neighbors,
+            k_bandwidth=k_bandwidth,
+            d_cdc=d_cdc,
+            gamma=gamma,
+            device="cuda" if torch.cuda.is_available() else "cpu",
+            debug=debug,
+            adaptive_k=adaptive_k,
+            min_bucket_size=min_bucket_size,
+            dataset_dirs=dataset_dirs
+        )
+
+        logger.info(f"CDC config hash: {preprocessor.config_hash}")
+
+        # Check if CDC caches already exist (unless force_recache)
+        if not force_recache:
+            all_cached = self._check_cdc_caches_exist(preprocessor.config_hash)
+            if all_cached:
+                logger.info("All CDC cache files found, skipping preprocessing")
+                return preprocessor.config_hash
             else:
-                logger.info(f"CDC cache found but invalid, will recompute")
+                logger.info("Some CDC cache files missing, will compute")
 
         # Only main process computes CDC
         is_main = accelerator is None or accelerator.is_main_process
         if not is_main:
             if accelerator is not None:
                 accelerator.wait_for_everyone()
-            return str(cdc_path)
+            return preprocessor.config_hash
 
-        logger.info("=" * 60)
         logger.info("Starting CDC-FM preprocessing")
         logger.info(f"Parameters: k={k_neighbors}, k_bw={k_bandwidth}, d_cdc={d_cdc}, gamma={gamma}")
-        logger.info("=" * 60)
-        # Initialize CDC preprocessor
-        # Initialize CDC preprocessor
-        try:
-            from library.cdc_fm import CDCPreprocessor
-        except ImportError as e:
-            logger.warning(
-                "FAISS not installed. CDC-FM preprocessing skipped. "
-                "Install with: pip install faiss-cpu (CPU) or faiss-gpu (GPU)"
-            )
-            return None
-
-        preprocessor = CDCPreprocessor(
-            k_neighbors=k_neighbors, k_bandwidth=k_bandwidth, d_cdc=d_cdc, gamma=gamma, device="cuda" if torch.cuda.is_available() else "cpu", debug=debug, adaptive_k=adaptive_k, min_bucket_size=min_bucket_size
-        )
 
         # Get caching strategy for loading latents
         from library.strategy_base import LatentsCachingStrategy
@@ -2789,45 +2805,61 @@ class DatasetGroup(torch.utils.data.ConcatDataset):
 
                 # Add to preprocessor (with unique global index across all datasets)
                 actual_global_idx = sum(len(d.image_data) for d in self.datasets[:dataset_idx]) + local_idx
-                preprocessor.add_latent(latent=latent, global_idx=actual_global_idx, shape=latent.shape, metadata={"image_key": info.image_key})
 
-        # Compute and save
+                # Get latents_npz_path - will be set whether caching to disk or memory
+                if info.latents_npz is None:
+                    # If not set, generate the path from the caching strategy
+                    info.latents_npz = caching_strategy.get_latents_npz_path(info.absolute_path, info.bucket_reso)
+
+                preprocessor.add_latent(
+                    latent=latent,
+                    global_idx=actual_global_idx,
+                    latents_npz_path=info.latents_npz,
+                    shape=latent.shape,
+                    metadata={"image_key": info.image_key}
+                )
+
+        # Compute and save individual CDC files
         logger.info(f"\nComputing CDC Γ_b matrices for {len(preprocessor.batcher)} samples...")
-        preprocessor.compute_all(save_path=cdc_path)
+        files_saved = preprocessor.compute_all()
+        logger.info(f"Saved {files_saved} CDC cache files")
 
         if accelerator is not None:
             accelerator.wait_for_everyone()
 
-        return str(cdc_path)
+        # Return config hash so training can initialize GammaBDataset with it
+        return preprocessor.config_hash
 
-    def _is_cdc_cache_valid(self, cdc_path: "pathlib.Path", k_neighbors: int, d_cdc: int, gamma: float) -> bool:
-        """Check if CDC cache has matching hyperparameters"""
-        try:
-            from safetensors import safe_open
+    def _check_cdc_caches_exist(self, config_hash: str) -> bool:
+        """
+        Check if CDC cache files exist for all latents in the dataset
 
-            with safe_open(str(cdc_path), framework="pt", device="cpu") as f:
-                cached_k = int(f.get_tensor("metadata/k_neighbors").item())
-                cached_d = int(f.get_tensor("metadata/d_cdc").item())
-                cached_gamma = float(f.get_tensor("metadata/gamma").item())
-                cached_num = int(f.get_tensor("metadata/num_samples").item())
+        Args:
+            config_hash: The config hash to use for CDC filename lookup
+        """
+        from pathlib import Path
 
-            expected_num = sum(len(d.image_data) for d in self.datasets)
+        missing_count = 0
+        total_count = 0
 
-            valid = cached_k == k_neighbors and cached_d == d_cdc and abs(cached_gamma - gamma) < 1e-6 and cached_num == expected_num
+        for dataset in self.datasets:
+            for info in dataset.image_data.values():
+                total_count += 1
+                if info.latents_npz is None:
+                    # If latents_npz not set, we can't check for CDC cache
+                    continue
 
-            if not valid:
-                logger.info(
-                    f"Cache mismatch: k={cached_k} (expected {k_neighbors}), "
-                    f"d_cdc={cached_d} (expected {d_cdc}), "
-                    f"gamma={cached_gamma} (expected {gamma}), "
-                    f"num={cached_num} (expected {expected_num})"
-                )
+                cdc_path = CDCPreprocessor.get_cdc_npz_path(info.latents_npz, config_hash)
+                if not Path(cdc_path).exists():
+                    missing_count += 1
 
-            return valid
-        except Exception as e:
-            logger.warning(f"Error validating CDC cache: {e}")
+        if missing_count > 0:
+            logger.info(f"Found {missing_count}/{total_count} missing CDC cache files")
             return False
 
+        logger.debug(f"All {total_count} CDC cache files exist")
+        return True
+
     def set_caching_mode(self, caching_mode):
         for dataset in self.datasets:
             dataset.set_caching_mode(caching_mode)
diff --git a/tests/library/test_cdc_preprocessor.py b/tests/library/test_cdc_preprocessor.py
index 17d159d7..63db6286 100644
--- a/tests/library/test_cdc_preprocessor.py
+++ b/tests/library/test_cdc_preprocessor.py
@@ -35,28 +35,38 @@ class TestCDCPreprocessorIntegration:
         # Add 10 small latents
         for i in range(10):
             latent = torch.randn(16, 4, 4, dtype=torch.float32)  # C, H, W
+            latents_npz_path = str(tmp_path / f"test_image_{i}_0004x0004_flux.npz")
             metadata = {'image_key': f'test_image_{i}'}
-            preprocessor.add_latent(latent=latent, global_idx=i, shape=latent.shape, metadata=metadata)
+            preprocessor.add_latent(
+                latent=latent,
+                global_idx=i,
+                latents_npz_path=latents_npz_path,
+                shape=latent.shape,
+                metadata=metadata
+            )
 
         # Compute and save
-        output_path = tmp_path / "test_gamma_b.safetensors"
-        result_path = preprocessor.compute_all(save_path=output_path)
+        files_saved = preprocessor.compute_all()
 
-        # Verify file was created
-        assert Path(result_path).exists()
+        # Verify files were created
+        assert files_saved == 10
 
-        # Verify structure
-        with safe_open(str(result_path), framework="pt", device="cpu") as f:
-            assert f.get_tensor("metadata/num_samples").item() == 10
-            assert f.get_tensor("metadata/k_neighbors").item() == 5
-            assert f.get_tensor("metadata/d_cdc").item() == 4
+        # Verify first CDC file structure
+        cdc_path = tmp_path / "test_image_0_0004x0004_flux_cdc.npz"
+        assert cdc_path.exists()
 
-            # Check first sample
-            eigvecs = f.get_tensor("eigenvectors/test_image_0")
-            eigvals = f.get_tensor("eigenvalues/test_image_0")
+        import numpy as np
+        data = np.load(cdc_path)
 
-            assert eigvecs.shape[0] == 4  # d_cdc
-            assert eigvals.shape[0] == 4  # d_cdc
+        assert data['k_neighbors'] == 5
+        assert data['d_cdc'] == 4
+
+        # Check eigenvectors and eigenvalues
+        eigvecs = data['eigenvectors']
+        eigvals = data['eigenvalues']
+
+        assert eigvecs.shape[0] == 4  # d_cdc
+        assert eigvals.shape[0] == 4  # d_cdc
 
     def test_preprocessor_with_different_shapes(self, tmp_path):
         """
@@ -69,27 +79,42 @@ class TestCDCPreprocessorIntegration:
         # Add 5 latents of shape (16, 4, 4)
         for i in range(5):
             latent = torch.randn(16, 4, 4, dtype=torch.float32)
+            latents_npz_path = str(tmp_path / f"test_image_{i}_0004x0004_flux.npz")
             metadata = {'image_key': f'test_image_{i}'}
-            preprocessor.add_latent(latent=latent, global_idx=i, shape=latent.shape, metadata=metadata)
+            preprocessor.add_latent(
+                latent=latent,
+                global_idx=i,
+                latents_npz_path=latents_npz_path,
+                shape=latent.shape,
+                metadata=metadata
+            )
 
         # Add 5 latents of different shape (16, 8, 8)
         for i in range(5, 10):
             latent = torch.randn(16, 8, 8, dtype=torch.float32)
+            latents_npz_path = str(tmp_path / f"test_image_{i}_0008x0008_flux.npz")
             metadata = {'image_key': f'test_image_{i}'}
-            preprocessor.add_latent(latent=latent, global_idx=i, shape=latent.shape, metadata=metadata)
+            preprocessor.add_latent(
+                latent=latent,
+                global_idx=i,
+                latents_npz_path=latents_npz_path,
+                shape=latent.shape,
+                metadata=metadata
+            )
 
         # Compute and save
-        output_path = tmp_path / "test_gamma_b_multi.safetensors"
-        result_path = preprocessor.compute_all(save_path=output_path)
+        files_saved = preprocessor.compute_all()
 
         # Verify both shape groups were processed
-        with safe_open(str(result_path), framework="pt", device="cpu") as f:
-            # Check shapes are stored
-            shape_0 = f.get_tensor("shapes/test_image_0")
-            shape_5 = f.get_tensor("shapes/test_image_5")
+        assert files_saved == 10
 
-            assert tuple(shape_0.tolist()) == (16, 4, 4)
-            assert tuple(shape_5.tolist()) == (16, 8, 8)
+        import numpy as np
+        # Check shapes are stored in individual files
+        data_0 = np.load(tmp_path / "test_image_0_0004x0004_flux_cdc.npz")
+        data_5 = np.load(tmp_path / "test_image_5_0008x0008_flux_cdc.npz")
+
+        assert tuple(data_0['shape']) == (16, 4, 4)
+        assert tuple(data_5['shape']) == (16, 8, 8)
 
 
 class TestDeviceConsistency:
@@ -107,19 +132,27 @@ class TestDeviceConsistency:
         )
 
         shape = (16, 32, 32)
+        latents_npz_paths = []
         for i in range(10):
             latent = torch.randn(*shape, dtype=torch.float32)
+            latents_npz_path = str(tmp_path / f"test_image_{i}_0032x0032_flux.npz")
+            latents_npz_paths.append(latents_npz_path)
             metadata = {'image_key': f'test_image_{i}'}
-            preprocessor.add_latent(latent=latent, global_idx=i, shape=shape, metadata=metadata)
+            preprocessor.add_latent(
+                latent=latent,
+                global_idx=i,
+                latents_npz_path=latents_npz_path,
+                shape=shape,
+                metadata=metadata
+            )
 
-        cache_path = tmp_path / "test_device.safetensors"
-        preprocessor.compute_all(save_path=cache_path)
+        preprocessor.compute_all()
 
-        dataset = GammaBDataset(gamma_b_path=cache_path, device="cpu")
+        dataset = GammaBDataset(device="cpu")
 
         noise = torch.randn(2, *shape, dtype=torch.float32, device="cpu")
         timesteps = torch.tensor([100.0, 200.0], dtype=torch.float32, device="cpu")
-        image_keys = ['test_image_0', 'test_image_1']
+        latents_npz_paths_batch = latents_npz_paths[:2]
 
         with caplog.at_level(logging.WARNING):
             caplog.clear()
@@ -128,7 +161,7 @@ class TestDeviceConsistency:
                 timesteps=timesteps,
                 num_timesteps=1000,
                 gamma_b_dataset=dataset,
-                image_keys=image_keys,
+                latents_npz_paths=latents_npz_paths_batch,
                 device="cpu"
             )
 
@@ -146,20 +179,28 @@ class TestDeviceConsistency:
         )
 
         shape = (16, 32, 32)
+        latents_npz_paths = []
         for i in range(10):
             latent = torch.randn(*shape, dtype=torch.float32)
+            latents_npz_path = str(tmp_path / f"test_image_{i}_0032x0032_flux.npz")
+            latents_npz_paths.append(latents_npz_path)
             metadata = {'image_key': f'test_image_{i}'}
-            preprocessor.add_latent(latent=latent, global_idx=i, shape=shape, metadata=metadata)
+            preprocessor.add_latent(
+                latent=latent,
+                global_idx=i,
+                latents_npz_path=latents_npz_path,
+                shape=shape,
+                metadata=metadata
+            )
 
-        cache_path = tmp_path / "test_device_mismatch.safetensors"
-        preprocessor.compute_all(save_path=cache_path)
+        preprocessor.compute_all()
 
-        dataset = GammaBDataset(gamma_b_path=cache_path, device="cpu")
+        dataset = GammaBDataset(device="cpu")
 
         # Create noise and timesteps
         noise = torch.randn(2, *shape, dtype=torch.float32, device="cpu", requires_grad=True)
         timesteps = torch.tensor([100.0, 200.0], dtype=torch.float32, device="cpu")
-        image_keys = ['test_image_0', 'test_image_1']
+        latents_npz_paths_batch = latents_npz_paths[:2]
 
         # Perform CDC transformation
         result = apply_cdc_noise_transformation(
@@ -167,7 +208,7 @@ class TestDeviceConsistency:
             timesteps=timesteps,
             num_timesteps=1000,
             gamma_b_dataset=dataset,
-            image_keys=image_keys,
+            latents_npz_paths=latents_npz_paths_batch,
             device="cpu"
         )
 
@@ -199,27 +240,34 @@ class TestCDCEndToEnd:
         )
 
         num_samples = 10
+        latents_npz_paths = []
         for i in range(num_samples):
             latent = torch.randn(16, 4, 4, dtype=torch.float32)
+            latents_npz_path = str(tmp_path / f"test_image_{i}_0004x0004_flux.npz")
+            latents_npz_paths.append(latents_npz_path)
             metadata = {'image_key': f'test_image_{i}'}
-            preprocessor.add_latent(latent=latent, global_idx=i, shape=latent.shape, metadata=metadata)
+            preprocessor.add_latent(
+                latent=latent,
+                global_idx=i,
+                latents_npz_path=latents_npz_path,
+                shape=latent.shape,
+                metadata=metadata
+            )
 
-        output_path = tmp_path / "cdc_gamma_b.safetensors"
-        cdc_path = preprocessor.compute_all(save_path=output_path)
+        files_saved = preprocessor.compute_all()
+        assert files_saved == num_samples
 
         # Step 2: Load with GammaBDataset
-        gamma_b_dataset = GammaBDataset(gamma_b_path=cdc_path, device="cpu")
-
-        assert gamma_b_dataset.num_samples == num_samples
+        gamma_b_dataset = GammaBDataset(device="cpu")
 
         # Step 3: Use in mock training scenario
         batch_size = 3
         batch_latents_flat = torch.randn(batch_size, 256)  # B, d (flattened 16*4*4=256)
         batch_t = torch.rand(batch_size)
-        image_keys = ['test_image_0', 'test_image_5', 'test_image_9']
+        latents_npz_paths_batch = [latents_npz_paths[0], latents_npz_paths[5], latents_npz_paths[9]]
 
         # Get Γ_b components
-        eigvecs, eigvals = gamma_b_dataset.get_gamma_b_sqrt(image_keys, device="cpu")
+        eigvecs, eigvals = gamma_b_dataset.get_gamma_b_sqrt(latents_npz_paths_batch, device="cpu")
 
         # Compute geometry-aware noise
         sigma_t_x = gamma_b_dataset.compute_sigma_t_x(eigvecs, eigvals, batch_latents_flat, batch_t)
diff --git a/train_network.py b/train_network.py
index 1fd0c8e5..88edcc10 100644
--- a/train_network.py
+++ b/train_network.py
@@ -687,9 +687,16 @@ class NetworkTrainer:
         if hasattr(args, "use_cdc_fm") and args.use_cdc_fm and self.cdc_cache_path is not None:
             from library.cdc_fm import GammaBDataset
 
-            logger.info(f"Loading CDC Γ_b dataset from {self.cdc_cache_path}")
+            # cdc_cache_path now contains the config hash
+            config_hash = self.cdc_cache_path if self.cdc_cache_path != "per_file" else None
+            if config_hash:
+                logger.info(f"CDC Γ_b dataset ready (hash: {config_hash})")
+            else:
+                logger.info("CDC Γ_b dataset ready (no hash, backward compatibility)")
+
             self.gamma_b_dataset = GammaBDataset(
-                gamma_b_path=self.cdc_cache_path, device="cuda" if torch.cuda.is_available() else "cpu"
+                device="cuda" if torch.cuda.is_available() else "cpu",
+                config_hash=config_hash
             )
         else:
             self.gamma_b_dataset = None