diff --git a/DENOISING_DIFFUSION/configs/base_config.yaml b/DENOISING_DIFFUSION/configs/base_config.yaml
new file mode 100644
index 0000000..c02a22f
--- /dev/null
+++ b/DENOISING_DIFFUSION/configs/base_config.yaml
@@ -0,0 +1,43 @@
+project:
+  name: exxa-denoising-diffusion
+  seed: 42
+
+data:
+  train_noisy_path: data/dirty.npy
+  train_clean_path: data/clean.npy
+  image_size: 64
+  patch_size: 64
+  patches_per_image: 4
+  val_split: 0.1
+  test_split: 0.1
+  num_workers: 2
+
+model:
+  in_channels: 1
+  out_channels: 1
+  base_channels: 64
+  channel_multipliers: [1, 2, 4, 8]
+  dropout: 0.1
+
+diffusion:
+  timesteps: 1000
+  beta_schedule: linear
+  beta_start: 0.0001
+  beta_end: 0.02
+
+training:
+  batch_size: 16
+  epochs: 100
+  learning_rate: 0.0002
+  weight_decay: 0.00001
+  grad_clip_norm: 1.0
+  use_amp: false
+  device: auto
+
+logging:
+  log_every_n_steps: 100
+  val_every_n_epochs: 1
+  save_every_n_epochs: 5
+  output_dir: experiments
+  use_wandb: true
+  wandb_project: exxa-denoising
diff --git a/DENOISING_DIFFUSION/src/training/__init__.py b/DENOISING_DIFFUSION/src/training/__init__.py
new file mode 100644
index 0000000..72dae0e
--- /dev/null
+++ b/DENOISING_DIFFUSION/src/training/__init__.py
@@ -0,0 +1,25 @@
+"""Training utilities for EXXA denoising diffusion models."""
+
+from .config import (
+    ProjectConfig,
+    DataConfig,
+    ModelConfig,
+    DiffusionConfig,
+    TrainingConfig,
+    LoggingConfig,
+    ProjectMetadata,
+    load_config,
+    save_config,
+)
+
+__all__ = [
+    "ProjectConfig",
+    "DataConfig",
+    "ModelConfig",
+    "DiffusionConfig",
+    "TrainingConfig",
+    "LoggingConfig",
+    "ProjectMetadata",
+    "load_config",
+    "save_config",
+]
diff --git a/DENOISING_DIFFUSION/src/training/config.py b/DENOISING_DIFFUSION/src/training/config.py
new file mode 100644
index 0000000..f18ca0b
--- /dev/null
+++ b/DENOISING_DIFFUSION/src/training/config.py
@@ -0,0 +1,164 @@
+"""Configuration schema and IO helpers for training and inference."""
+
+from __future__ import annotations
+
+from dataclasses import asdict, dataclass, field
+from pathlib import Path
+from typing import Any, Mapping
+import copy
+import json
+
+import yaml
+
+
+@dataclass
+class ProjectMetadata:
+    name: str = "exxa-denoising-diffusion"
+    seed: int = 42
+
+
+@dataclass
+class DataConfig:
+    train_noisy_path: str = "data/dirty.npy"
+    train_clean_path: str = "data/clean.npy"
+    image_size: int = 64
+    patch_size: int = 64
+    patches_per_image: int = 4
+    val_split: float = 0.1
+    test_split: float = 0.1
+    num_workers: int = 2
+
+
+@dataclass
+class ModelConfig:
+    in_channels: int = 1
+    out_channels: int = 1
+    base_channels: int = 64
+    channel_multipliers: list[int] | tuple[int, ...] = (1, 2, 4, 8)
+    dropout: float = 0.1
+
+
+@dataclass
+class DiffusionConfig:
+    timesteps: int = 1000
+    beta_schedule: str = "linear"
+    beta_start: float = 1e-4
+    beta_end: float = 2e-2
+
+
+@dataclass
+class TrainingConfig:
+    batch_size: int = 16
+    epochs: int = 100
+    learning_rate: float = 2e-4
+    weight_decay: float = 1e-5
+    grad_clip_norm: float = 1.0
+    use_amp: bool = False
+    device: str = "auto"
+
+
+@dataclass
+class LoggingConfig:
+    log_every_n_steps: int = 100
+    val_every_n_epochs: int = 1
+    save_every_n_epochs: int = 5
+    output_dir: str = "experiments"
+    use_wandb: bool = True
+    wandb_project: str = "exxa-denoising"
+
+
+@dataclass
+class ProjectConfig:
+    project: ProjectMetadata = field(default_factory=ProjectMetadata)
+    data: DataConfig = field(default_factory=DataConfig)
+    model: ModelConfig = field(default_factory=ModelConfig)
+    diffusion: DiffusionConfig = field(default_factory=DiffusionConfig)
+    training: TrainingConfig = field(default_factory=TrainingConfig)
+    logging: LoggingConfig = field(default_factory=LoggingConfig)
+
+    def to_dict(self) -> dict[str, Any]:
+        return asdict(self)
+
+    def validate(self) -> None:
+        if self.data.image_size <= 0:
+            raise ValueError("data.image_size must be > 0")
+        if self.training.batch_size <= 0:
+            raise ValueError("training.batch_size must be > 0")
+        if self.training.learning_rate <= 0:
+            raise ValueError("training.learning_rate must be > 0")
+        if self.diffusion.timesteps <= 0:
+            raise ValueError("diffusion.timesteps must be > 0")
+        if self.diffusion.beta_schedule not in {"linear", "cosine"}:
+            raise ValueError("diffusion.beta_schedule must be one of: linear, cosine")
+        split_total = self.data.val_split + self.data.test_split
+        if not (0.0 <= self.data.val_split < 1.0 and 0.0 <= self.data.test_split < 1.0):
+            raise ValueError("data.val_split and data.test_split must be in [0, 1)")
+        if split_total >= 1.0:
+            raise ValueError("data.val_split + data.test_split must be < 1")
+
+    @classmethod
+    def from_dict(cls, payload: Mapping[str, Any]) -> "ProjectConfig":
+        data = copy.deepcopy(payload)
+        cfg = cls(
+            project=ProjectMetadata(**data.get("project", {})),
+            data=DataConfig(**data.get("data", {})),
+            model=ModelConfig(**data.get("model", {})),
+            diffusion=DiffusionConfig(**data.get("diffusion", {})),
+            training=TrainingConfig(**data.get("training", {})),
+            logging=LoggingConfig(**data.get("logging", {})),
+        )
+        cfg.validate()
+        return cfg
+
+
+def _deep_merge(base: dict[str, Any], updates: Mapping[str, Any]) -> dict[str, Any]:
+    merged = copy.deepcopy(base)
+    for key, value in updates.items():
+        if key in merged and isinstance(merged[key], dict) and isinstance(value, Mapping):
+            merged[key] = _deep_merge(merged[key], value)
+        else:
+            merged[key] = value
+    return merged
+
+
+def _read_config_file(path: Path) -> dict[str, Any]:
+    suffix = path.suffix.lower()
+    with path.open("r", encoding="utf-8") as f:
+        if suffix in {".yaml", ".yml"}:
+            data = yaml.safe_load(f) or {}
+        elif suffix == ".json":
+            data = json.load(f)
+        else:
+            raise ValueError(f"Unsupported config extension: {suffix}")
+    if not isinstance(data, dict):
+        raise ValueError("Config root must be a mapping")
+    return data
+
+
+def load_config(
+    config_path: str | Path | None = None,
+    overrides: Mapping[str, Any] | None = None,
+) -> ProjectConfig:
+    """Load project config from file and optional nested overrides."""
+    if config_path is None:
+        config_path = Path(__file__).resolve().parents[2] / "configs" / "base_config.yaml"
+    path = Path(config_path)
+    payload = _read_config_file(path)
+    if overrides:
+        payload = _deep_merge(payload, overrides)
+    return ProjectConfig.from_dict(payload)
+
+
+def save_config(config: ProjectConfig, config_path: str | Path) -> None:
+    """Save config to YAML or JSON based on destination extension."""
+    path = Path(config_path)
+    path.parent.mkdir(parents=True, exist_ok=True)
+    data = config.to_dict()
+
+    with path.open("w", encoding="utf-8") as f:
+        if path.suffix.lower() in {".yaml", ".yml"}:
+            yaml.safe_dump(data, f, sort_keys=False)
+        elif path.suffix.lower() == ".json":
+            json.dump(data, f, indent=2)
+        else:
+            raise ValueError(f"Unsupported config extension: {path.suffix}")
diff --git a/DENOISING_DIFFUSION/tests/test_config.py b/DENOISING_DIFFUSION/tests/test_config.py
new file mode 100644
index 0000000..c236815
--- /dev/null
+++ b/DENOISING_DIFFUSION/tests/test_config.py
@@ -0,0 +1,68 @@
+"""Unit tests for training configuration schema and IO."""
+
+from pathlib import Path
+import json
+import sys
+
+import pytest
+
+sys.path.insert(0, str(Path(__file__).parent.parent / "src"))
+
+from training.config import ProjectConfig, load_config, save_config
+
+
+def test_load_default_config() -> None:
+    cfg = load_config()
+    assert isinstance(cfg, ProjectConfig)
+    assert cfg.project.name == "exxa-denoising-diffusion"
+    assert cfg.diffusion.timesteps == 1000
+
+
+def test_load_with_overrides() -> None:
+    cfg = load_config(
+        overrides={
+            "training": {"batch_size": 8},
+            "diffusion": {"beta_schedule": "cosine"},
+        }
+    )
+    assert cfg.training.batch_size == 8
+    assert cfg.diffusion.beta_schedule == "cosine"
+
+
+def test_load_json_config(tmp_path: Path) -> None:
+    config_path = tmp_path / "custom_config.json"
+    payload = {
+        "project": {"name": "json-test", "seed": 7},
+        "data": {"image_size": 128},
+        "model": {"base_channels": 32},
+        "diffusion": {"timesteps": 200, "beta_schedule": "linear"},
+        "training": {"batch_size": 4, "learning_rate": 1e-4},
+        "logging": {"use_wandb": False},
+    }
+    config_path.write_text(json.dumps(payload), encoding="utf-8")
+
+    cfg = load_config(config_path)
+    assert cfg.project.name == "json-test"
+    assert cfg.data.image_size == 128
+    assert cfg.training.batch_size == 4
+
+
+def test_save_and_reload_yaml(tmp_path: Path) -> None:
+    cfg = load_config(overrides={"project": {"name": "saved-test"}})
+    out_path = tmp_path / "saved_config.yaml"
+
+    save_config(cfg, out_path)
+    loaded = load_config(out_path)
+
+    assert loaded.project.name == "saved-test"
+    assert loaded.model.base_channels == cfg.model.base_channels
+
+
+def test_validation_rejects_bad_split() -> None:
+    with pytest.raises(ValueError, match=r"val_split \+ data.test_split"):
+        load_config(overrides={"data": {"val_split": 0.7, "test_split": 0.4}})
+
+
+def test_validation_rejects_bad_schedule() -> None:
+    with pytest.raises(ValueError, match="beta_schedule"):
+        load_config(overrides={"diffusion": {"beta_schedule": "invalid"}})