diff --git a/mmgen/core/evaluation/evaluation.py b/mmgen/core/evaluation/evaluation.py
index b5b32bce6..e38170c11 100644
--- a/mmgen/core/evaluation/evaluation.py
+++ b/mmgen/core/evaluation/evaluation.py
@@ -5,6 +5,7 @@
 from copy import deepcopy
 
 import mmcv
+import numpy as np
 import torch
 import torch.distributed as dist
 from mmcv.runner import get_dist_info
@@ -62,6 +63,89 @@ def make_vanilla_dataloader(img_path, batch_size, dist=False):
     return dataloader
 
 
+def make_npz_dataloader(npz_path, batch_size, dist=False):
+    pipeline = [
+        # permute the color channel. Because in npz_dataloader's pipeline, we
+        # direct load image in RGB order from npz file and we must convert it
+        # to BGR by setting ``to_rgb=True``.
+        dict(
+            type='Normalize',
+            keys=['real_img'],
+            mean=[127.5] * 3,
+            std=[127.5] * 3,
+            to_rgb=True),
+        dict(type='ImageToTensor', keys=['real_img']),
+        dict(type='Collect', keys=['real_img'])
+    ]
+
+    dataset = build_dataset(
+        dict(type='FileDataset', file_path=npz_path, pipeline=pipeline))
+    dataloader = build_dataloader(
+        dataset,
+        samples_per_gpu=batch_size,
+        workers_per_gpu=0,
+        dist=dist,
+        shuffle=True)
+    return dataloader
+
+
+def parse_npz_file_name(npz_file):
+    """Parse the basic information from the give npz file.
+    Args:
+        npz_file (str): The file name of the npz file.
+
+    Returns:
+        tuple(int): A tuple of (num_samples, H, W, num_channels).
+    """
+    # remove 'samples_' (8) and '.npz' (4)
+    num_samples, H, W, num_channles = npz_file[8:-4].split('x')
+    assert num_samples.isdigit()
+    assert H.isdigit()
+    assert W.isdigit()
+    assert num_channles.isdigit()
+    return int(num_samples), int(H), int(W), int(num_channles)
+
+
+def parse_npz_folder(npz_folder_path):
+    """Parse the npz files under the given folder.
+    Args:
+        npz_folder_path: The folder contains npz file.
+
+    Returns:
+        tuple(list, int): A tuple contains a list valid npz files' names and
+            a int of existing image numbers.
+    """
+
+    npz_files = [
+        f for f in list(mmcv.scandir(npz_folder_path, suffix=('.npz')))
+        if 'samples' in f
+    ]
+    valid_npz_files = []
+    img_shape = None
+    num_exist = 0
+    for npz_file in npz_files:
+        try:
+            n_samples, H, W, n_channels = parse_npz_file_name(npz_file)
+
+            # shape checking
+            if img_shape is None:
+                img_shape = (H, W, n_channels)
+            else:
+                if img_shape != (H, W, n_channels):
+                    raise ValueError(
+                        'Image shape conflicting under sample path:'
+                        f'\'{npz_folder_path}\'. Find {img_shape} vs. '
+                        f'{(H, W, n_channels)}.')
+
+            valid_npz_files.append(npz_file)
+            num_exist += n_samples
+        except AssertionError:
+            mmcv.print_log(
+                f'Find npz file \'{npz_file}\' does not conform to the '
+                'standard naming convention.', 'mmgen')
+    return valid_npz_files, num_exist
+
+
 @torch.no_grad()
 def offline_evaluation(model,
                        data_loader,
@@ -70,6 +154,7 @@ def offline_evaluation(model,
                        basic_table_info,
                        batch_size,
                        samples_path=None,
+                       save_npz=False,
                        **kwargs):
     """Evaluate model in offline mode.
 
@@ -87,6 +172,10 @@ def offline_evaluation(model,
         samples_path (str): Used to save generated images. If it's none, we'll
             give it a default directory and delete it after finishing the
             evaluation. Default to None.
+        save_npz (bool, optional): Whether save the generated images to a npz
+            file named 'samples_{NUM_IMAGES}x{H}x{W}x{NUM_CHANNELS}.npz' If
+            true, dataset will be build upon npz file instead of image files.
+            Defaults to True.
         kwargs (dict): Other arguments.
     """
     # eval special and recon metric online only
@@ -111,11 +200,14 @@ def offline_evaluation(model,
         os.makedirs(samples_path)
         delete_samples_path = True
 
-    # sample images
-    num_exist = len(
-        list(
-            mmcv.scandir(
-                samples_path, suffix=('.jpg', '.png', '.jpeg', '.JPEG'))))
+    # check existing images
+    if save_npz:
+        npz_file_exist, num_exist = parse_npz_folder(samples_path)
+    else:
+        num_exist = len(
+            list(
+                mmcv.scandir(
+                    samples_path, suffix=('.jpg', '.png', '.jpeg', '.JPEG'))))
     if basic_table_info['num_samples'] > 0:
         max_num_images = basic_table_info['num_samples']
     else:
@@ -128,6 +220,7 @@ def offline_evaluation(model,
         # define mmcv progress bar
         pbar = mmcv.ProgressBar(num_needed)
 
+    fake_img_list = []
     # if no images, `num_needed` should be zero
     total_batch_size = batch_size * ws
     for begin in range(0, num_needed, total_batch_size):
@@ -163,8 +256,66 @@ def offline_evaluation(model,
                 images = fakes[i:i + 1]
                 images = ((images + 1) / 2)
                 images = images.clamp_(0, 1)
-                image_name = str(num_exist + begin + i) + '.png'
-                save_image(images, os.path.join(samples_path, image_name))
+                if save_npz:
+                    # permute to [H, W, chn] and rescale to [0, 255]
+                    fake_img_list.append(
+                        images.permute(0, 2, 3, 1).cpu().numpy() * 255)
+                else:
+                    image_name = str(num_exist + begin + i) + '.png'
+                    save_image(images, os.path.join(samples_path, image_name))
+
+    if save_npz:
+        # only one npz file and fake_img_list is empty --> do not need to save
+        if len(npz_file_exist) == 1 and len(fake_img_list) == 0:
+            npz_path = os.path.join(samples_path, npz_file_exist[0])
+            if rank == 0:
+                mmcv.print_log(
+                    f'Existing npz file \'{npz_path}\' has already met '
+                    'requirements.', 'mmgen')
+        else:
+            # load from locl file and merge to one
+            if rank == 0:
+                fake_img_exist_list = []
+                for exist_npz_file in npz_file_exist:
+                    fake_imgs_ = np.load(
+                        os.path.join(samples_path, exist_npz_file))['real_img']
+                    fake_img_exist_list.append(fake_imgs_)
+
+                # merge fake_img_exist_list and fake_img_list
+                fake_imgs = np.concatenate(
+                    fake_img_exist_list + fake_img_list, axis=0)
+                num_imgs, H, W, num_channels = fake_imgs.shape
+
+                npz_path = os.path.join(
+                    samples_path,
+                    f'samples_{num_imgs}x{H}x{W}x{num_channels}.npz')
+
+                # save new npz file -->
+                # set key as ``real_img`` to align with vanilla dataset
+                np.savez(npz_path, real_img=fake_imgs)
+                mmcv.print_log(f'Save new npz_file to \'{npz_path}\'.',
+                               'mmgen')
+
+                # delete old npz files
+                for npz_file in npz_file_exist:
+                    os.remove(os.path.join(samples_path, npz_file))
+                    mmcv.print_log(
+                        'Remove useless npz file '
+                        f'\'{os.path.join(samples_path, npz_file)}\'.',
+                        'mmgen')
+
+            # waiting for rank-0 to save the new npz file
+            if ws > 1:
+                dist.barrier()
+            # get npz_path.
+            # We have delete useless npz files then there should only one
+            # file under the sample_path. Check and directly load it!
+            npz_files = [
+                f for f in list(mmcv.scandir(samples_path, suffix=('.npz')))
+                if 'samples' in f
+            ]
+            assert len(npz_files) == 1
+            npz_path = os.path.join(samples_path, npz_files[0])
 
     if num_needed > 0 and rank == 0:
         sys.stdout.write('\n')
@@ -175,8 +326,13 @@ def offline_evaluation(model,
 
     # empty cache to release GPU memory
     torch.cuda.empty_cache()
-    fake_dataloader = make_vanilla_dataloader(
-        samples_path, batch_size, dist=ws > 1)
+    if save_npz:
+        fake_dataloader = make_npz_dataloader(
+            npz_path, batch_size, dist=ws > 1)
+    else:
+        fake_dataloader = make_vanilla_dataloader(
+            samples_path, batch_size, dist=ws > 1)
+
     for metric in metrics:
         mmcv.print_log(f'Evaluate with {metric.name} metric.', 'mmgen')
         metric.prepare()
diff --git a/mmgen/datasets/__init__.py b/mmgen/datasets/__init__.py
index 3a35a4aba..83d458769 100644
--- a/mmgen/datasets/__init__.py
+++ b/mmgen/datasets/__init__.py
@@ -1,6 +1,7 @@
 # Copyright (c) OpenMMLab. All rights reserved.
 from .builder import build_dataloader, build_dataset
 from .dataset_wrappers import RepeatDataset
+from .file_dataset import FileDataset
 from .grow_scale_image_dataset import GrowScaleImgDataset
 from .paired_image_dataset import PairedImageDataset
 from .pipelines import (Collect, Compose, Flip, ImageToTensor,
@@ -16,5 +17,5 @@
     'DistributedSampler', 'UnconditionalImageDataset', 'Compose', 'ToTensor',
     'ImageToTensor', 'Collect', 'Flip', 'Resize', 'RepeatDataset', 'Normalize',
     'GrowScaleImgDataset', 'SinGANDataset', 'PairedImageDataset',
-    'UnpairedImageDataset', 'QuickTestImageDataset'
+    'UnpairedImageDataset', 'QuickTestImageDataset', 'FileDataset'
 ]
diff --git a/mmgen/datasets/file_dataset.py b/mmgen/datasets/file_dataset.py
new file mode 100644
index 000000000..761ec94ea
--- /dev/null
+++ b/mmgen/datasets/file_dataset.py
@@ -0,0 +1,126 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import mmcv
+import numpy as np
+from torch.utils.data import Dataset
+
+from .builder import DATASETS
+from .pipelines import Compose
+
+
+@DATASETS.register_module()
+class FileDataset(Dataset):
+    """Uncoditional file Dataset.
+
+    This dataset load data information from files for training GANs. Given
+    the path of a file, we will load all information in the file. The
+    transformation on data is defined by the pipeline. Please ensure that
+    ``LoadImageFromFile`` is not in your pipeline configs because we directly
+    get images in ``np.ndarray`` from the given file.
+
+    Args:
+        file_path (str): Path of the file.
+        img_keys (str): Key of the images in npz file.
+        pipeline (list[dict | callable]): A sequence of data transforms.
+        test_mode (bool, optional): If True, the dataset will work in test
+            mode. Otherwise, in train mode. Default to False.
+        npz_keys (str | list[str], optional): Key of the images to load in the
+            npz file. Must with the input file is as npz file.
+    """
+
+    _VALID_FILE_SUFFIX = ('.npz')
+
+    def __init__(self, file_path, pipeline, test_mode=False):
+        super().__init__()
+        assert any([
+            file_path.endswith(suffix) for suffix in self._VALID_FILE_SUFFIX
+        ]), (f'We only support \'{self._VALID_FILE_SUFFIX}\' in this dataset, '
+             f'but receive {file_path}.')
+
+        self.file_path = file_path
+        self.pipeline = Compose(pipeline)
+        self.test_mode = test_mode
+        self.load_annotations()
+
+        # print basic dataset information to check the validity
+        mmcv.print_log(repr(self), 'mmgen')
+
+    def load_annotations(self):
+        """Load annotations."""
+        if self.file_path.endswith('.npz'):
+            data_info, data_length = self._load_annotations_from_npz()
+            data_fetch_fn = self._npz_data_fetch_fn
+
+        self.data_infos = data_info
+        self.data_fetch_fn = data_fetch_fn
+        self.data_length = data_length
+
+    def _load_annotations_from_npz(self):
+        """Load annotations from npz file and check number of samples are
+        consistent  among all items.
+
+        Returns:
+            tuple: dict and int
+        """
+        npz_file = np.load(self.file_path, mmap_mode='r')
+        data_info_dict = dict()
+        npz_keys = list(npz_file.keys())
+
+        # checnk num samples
+        num_samples = None
+        for k in npz_keys:
+            data_info_dict[k] = npz_file[k]
+            # check number of samples
+            if num_samples is None:
+                num_samples = npz_file[k].shape[0]
+            else:
+                assert num_samples == npz_file[k].shape[0]
+        return data_info_dict, num_samples
+
+    @staticmethod
+    def _npz_data_fetch_fn(data_infos, idx):
+        """Fetch data from npz file by idx and package them to a dict.
+
+        Args:
+            data_infos (array, tuple, dict): Data infos in the npz file.
+            idx (int): Index of current batch.
+
+        Returns:
+            dict: Data infos of the given idx.
+        """
+        data_dict = dict()
+        for k in data_infos.keys():
+            if data_infos[k][idx].shape == ():
+                v = np.array([data_infos[k][idx]])
+            else:
+                v = data_infos[k][idx]
+            data_dict[k] = v
+        return data_dict
+
+    def prepare_data(self, idx, data_fetch_fn=None):
+        """Prepare data.
+
+        Args:
+            idx (int): Index of current batch.
+            data_fetch_fn (callable): Function to fetch data.
+
+        Returns:
+            dict: Prepared training data batch.
+        """
+        if data_fetch_fn is None:
+            data = self.data_infos[idx]
+        else:
+            data = data_fetch_fn(self.data_infos, idx)
+        return self.pipeline(data)
+
+    def __len__(self):
+        return self.data_length
+
+    def __getitem__(self, idx):
+        return self.prepare_data(idx, self.data_fetch_fn)
+
+    def __repr__(self):
+        dataset_name = self.__class__
+        file_path = self.file_path
+        num_imgs = len(self)
+        return (f'dataset_name: {dataset_name}, total {num_imgs} images in '
+                f'file_path: {file_path}')
diff --git a/mmgen/datasets/pipelines/formatting.py b/mmgen/datasets/pipelines/formatting.py
index 37a52d39d..93ce1125d 100644
--- a/mmgen/datasets/pipelines/formatting.py
+++ b/mmgen/datasets/pipelines/formatting.py
@@ -129,8 +129,9 @@ def __call__(self, results):
         """
         data = {}
         img_meta = {}
-        for key in self.meta_keys:
-            img_meta[key] = results[key]
+        if self.meta_keys is not None:
+            for key in self.meta_keys:
+                img_meta[key] = results[key]
         data['meta'] = DC(img_meta, cpu_only=True)
         for key in self.keys:
             data[key] = results[key]
diff --git a/tests/data/file/test.npz b/tests/data/file/test.npz
new file mode 100644
index 000000000..7210664f2
Binary files /dev/null and b/tests/data/file/test.npz differ
diff --git a/tests/test_datasets/test_file_image_dataset.py b/tests/test_datasets/test_file_image_dataset.py
new file mode 100644
index 000000000..e87c467ef
--- /dev/null
+++ b/tests/test_datasets/test_file_image_dataset.py
@@ -0,0 +1,31 @@
+import os.path as osp
+
+from mmgen.datasets import FileDataset
+
+
+class TestFileDataset(object):
+
+    @classmethod
+    def setup_class(cls):
+        cls.file_path = osp.join(
+            osp.dirname(__file__), '..', 'data/file/test.npz')
+        cls.default_pipeline = [
+            dict(type='Resize', scale=(32, 32), keys=['fake_img']),
+            dict(type='ToTensor', keys=['label']),
+            dict(type='ImageToTensor', keys=['fake_img']),
+            dict(type='Collect', keys=['fake_img', 'label'])
+        ]
+
+    def test_unconditional_imgs_dataset(self):
+        dataset = FileDataset(self.file_path, pipeline=self.default_pipeline)
+
+        assert len(dataset) == 2
+        data_dict = dataset[0]
+        img = data_dict['fake_img']
+        lab = data_dict['label']
+        assert img.shape == (3, 32, 32)
+        assert lab == 1
+        print(repr(dataset))
+        assert repr(dataset) == (
+            f'dataset_name: {dataset.__class__}, '
+            f'total {2} images in file_path: {self.file_path}')
diff --git a/tests/test_datasets/test_pipelines/test_formatting.py b/tests/test_datasets/test_pipelines/test_formatting.py
index 4ccb91005..226bf3c26 100644
--- a/tests/test_datasets/test_pipelines/test_formatting.py
+++ b/tests/test_datasets/test_pipelines/test_formatting.py
@@ -90,7 +90,6 @@ def test_collect():
     collect = Collect(keys, meta_keys=meta_keys)
     results = collect(inputs)
     assert set(list(results.keys())) == set(['img', 'label', 'meta'])
-    inputs.pop('img')
     assert set(results['meta'].data.keys()) == set(meta_keys)
     for key in results['meta'].data:
         assert results['meta'].data[key] == inputs[key]
@@ -98,3 +97,9 @@ def test_collect():
     assert repr(collect) == (
         collect.__class__.__name__ +
         f'(keys={keys}, meta_keys={collect.meta_keys})')
+
+    # test meta is None
+    collect = Collect(keys)
+    results = collect(inputs)
+    print(results['meta'].data)
+    assert results['meta'].data == {}
diff --git a/tools/evaluation.py b/tools/evaluation.py
index 731c623a8..4b5a98a03 100644
--- a/tools/evaluation.py
+++ b/tools/evaluation.py
@@ -50,6 +50,12 @@ def parse_args():
         default=None,
         help='path to store images. If not given, remove it after evaluation\
              finished')
+    parser.add_argument(
+        '--save-npz',
+        action='store_true',
+        help=('whether to save generated images to a npz file named '
+              '\'NUM_IMAGES.npz\'. The npz file will be saved at '
+              '`samples-path`. (only work in offline mode)'))
     parser.add_argument(
         '--sample-model',
         type=str,
@@ -113,8 +119,6 @@ def main():
         init_dist(args.launcher, **cfg.dist_params)
         rank, world_size = get_dist_info()
         cfg.gpu_ids = range(world_size)
-        assert args.online or world_size == 1, (
-            'We only support online mode for distrbuted evaluation.')
 
     dirname = os.path.dirname(args.checkpoint)
     ckpt = os.path.basename(args.checkpoint)
@@ -218,7 +222,7 @@ def main():
     else:
         offline_evaluation(model, data_loader, metrics, logger,
                            basic_table_info, args.batch_size,
-                           args.samples_path, **args.sample_cfg)
+                           args.samples_path, args.save_npz, **args.sample_cfg)
 
 
 if __name__ == '__main__':