diff --git a/README.md b/README.md
index 5061bd1547..a6ba944a9e 100644
--- a/README.md
+++ b/README.md
@@ -34,6 +34,7 @@ import numpy as np
 
 generator = torch.Generator()
 generator = generator.manual_seed(6694729458485568)
+torch_device = "cuda" if torch.cuda.is_available() else "cpu"
 
 # 1. Load models
 scheduler = GaussianDDPMScheduler.from_config("fusing/ddpm-lsun-church")
diff --git a/examples/sample_loop.py b/examples/sample_loop.py
deleted file mode 100755
index d8134a6bf3..0000000000
--- a/examples/sample_loop.py
+++ /dev/null
@@ -1,157 +0,0 @@
-#!/usr/bin/env python3
-from diffusers import UNetModel, GaussianDDPMScheduler
-import torch
-import torch.nn.functional as F
-import numpy as np
-import PIL.Image
-import tqdm
-
-#torch_device = "cuda"
-#
-#unet = UNetModel.from_pretrained("/home/patrick/ddpm-lsun-church")
-#unet.to(torch_device)
-#
-#TIME_STEPS = 10
-#
-#scheduler = GaussianDDPMScheduler.from_config("/home/patrick/ddpm-lsun-church", timesteps=TIME_STEPS)
-#
-#diffusion_config = {
-#    "beta_start": 0.0001,
-#    "beta_end": 0.02,
-#    "num_diffusion_timesteps": TIME_STEPS,
-#}
-#
-# 2. Do one denoising step with model
-#batch_size, num_channels, height, width = 1, 3, 256, 256
-#
-#torch.manual_seed(0)
-#noise_image = torch.randn(batch_size, num_channels, height, width, device="cuda")
-#
-#
-# Helper
-#def noise_like(shape, device, repeat=False):
-#    def repeat_noise():
-#        return torch.randn((1, *shape[1:]), device=device).repeat(shape[0], *((1,) * (len(shape) - 1)))
-#
-#    def noise():
-#        return torch.randn(shape, device=device)
-#
-#    return repeat_noise() if repeat else noise()
-#
-#
-#betas = np.linspace(diffusion_config["beta_start"], diffusion_config["beta_end"], diffusion_config["num_diffusion_timesteps"], dtype=np.float64)
-#betas = torch.tensor(betas, device=torch_device)
-#alphas = 1.0 - betas
-#
-#alphas_cumprod = torch.cumprod(alphas, axis=0)
-#alphas_cumprod_prev = F.pad(alphas_cumprod[:-1], (1, 0), value=1.0)
-#
-#posterior_mean_coef1 = betas * torch.sqrt(alphas_cumprod_prev) / (1.0 - alphas_cumprod)
-#posterior_mean_coef2 = (1.0 - alphas_cumprod_prev) * torch.sqrt(alphas) / (1.0 - alphas_cumprod)
-#
-#posterior_variance = betas * (1.0 - alphas_cumprod_prev) / (1.0 - alphas_cumprod)
-#posterior_log_variance_clipped = torch.log(posterior_variance.clamp(min=1e-20))
-#
-#
-#sqrt_recip_alphas_cumprod = torch.sqrt(1.0 / alphas_cumprod)
-#sqrt_recipm1_alphas_cumprod = torch.sqrt(1.0 / alphas_cumprod - 1)
-#
-#
-#noise_coeff = (1 - alphas) / torch.sqrt(1 - alphas_cumprod)
-#coeff = 1 / torch.sqrt(alphas)
-
-
-def real_fn():
-    # Compare the following to Algorithm 2 Sampling of paper: https://arxiv.org/pdf/2006.11239.pdf
-    # 1: x_t ~ N(0,1)
-    x_t = noise_image
-    # 2: for t = T, ...., 1 do
-    for i in reversed(range(TIME_STEPS)):
-        t = torch.tensor([i]).to(torch_device)
-        # 3: z ~ N(0, 1)
-        noise = noise_like(x_t.shape, torch_device)
-
-        # 4:  √1αtxt − √1−αt1−α¯tθ(xt, t) + σtz
-        # ------------------------- MODEL ------------------------------------#
-        with torch.no_grad():
-            pred_noise = unet(x_t, t)  # pred epsilon_theta
-
-    #    pred_x = sqrt_recip_alphas_cumprod[t] * x_t - sqrt_recipm1_alphas_cumprod[t] * pred_noise
-    #    pred_x.clamp_(-1.0, 1.0)
-        # pred mean
-    #    posterior_mean = posterior_mean_coef1[t] * pred_x + posterior_mean_coef2[t] * x_t
-        # --------------------------------------------------------------------#
-
-        posterior_mean = coeff[t] * (x_t - noise_coeff[t] * pred_noise)
-
-        # ------------------------- Variance Scheduler -----------------------#
-        # pred variance
-        posterior_log_variance = posterior_log_variance_clipped[t]
-
-        b, *_, device = *x_t.shape, x_t.device
-        nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x_t.shape) - 1)))
-        posterior_variance = nonzero_mask * (0.5 * posterior_log_variance).exp()
-        # --------------------------------------------------------------------#
-
-        x_t_1 = (posterior_mean + posterior_variance * noise).to(torch.float32)
-        x_t = x_t_1
-
-        print(x_t.abs().sum())
-
-
-def post_process_to_image(x_t):
-    image = x_t.cpu().permute(0, 2, 3, 1)
-    image = (image + 1.0) * 127.5
-    image = image.numpy().astype(np.uint8)
-
-    return PIL.Image.fromarray(image[0])
-
-
-from pytorch_diffusion import Diffusion
-
-#diffusion = Diffusion.from_pretrained("lsun_church")
-#samples = diffusion.denoise(1)
-#
-#image = post_process_to_image(samples)
-#image.save("check.png")
-#import ipdb; ipdb.set_trace()
-
-
-device = "cuda"
-scheduler = GaussianDDPMScheduler.from_config("/home/patrick/ddpm-lsun-church", timesteps=10)
-
-import ipdb; ipdb.set_trace()
-
-model = UNetModel.from_pretrained("/home/patrick/ddpm-lsun-church").to(device)
-
-
-torch.manual_seed(0)
-next_image = scheduler.sample_noise((1, model.in_channels, model.resolution, model.resolution), device=device)
-
-for t in tqdm.tqdm(reversed(range(len(scheduler))), total=len(scheduler)):
-    # define coefficients for time step t
-    clip_image_coeff = 1 / torch.sqrt(scheduler.get_alpha_prod(t))
-    clip_noise_coeff = torch.sqrt(1 / scheduler.get_alpha_prod(t) - 1)
-    image_coeff = (1 - scheduler.get_alpha_prod(t - 1)) * torch.sqrt(scheduler.get_alpha(t)) / (1 - scheduler.get_alpha_prod(t))
-    clip_coeff = torch.sqrt(scheduler.get_alpha_prod(t - 1)) * scheduler.get_beta(t) / (1 - scheduler.get_alpha_prod(t))
-
-    # predict noise residual
-    with torch.no_grad():
-        noise_residual = model(next_image, t)
-
-    # compute prev image from noise
-    pred_mean = clip_image_coeff * next_image - clip_noise_coeff * noise_residual
-    pred_mean = torch.clamp(pred_mean, -1, 1)
-    image = clip_coeff * pred_mean + image_coeff * next_image
-
-    # sample variance
-    variance = scheduler.sample_variance(t, image.shape, device=device)
-
-    # sample previous image
-    sampled_image = image + variance
-
-    next_image = sampled_image
-
-
-image = post_process_to_image(next_image)
-image.save("example_new.png")
diff --git a/models/vision/ddpm/example.py b/models/vision/ddpm/example.py
index 2ba753c385..c086407490 100755
--- a/models/vision/ddpm/example.py
+++ b/models/vision/ddpm/example.py
@@ -1,20 +1,23 @@
 #!/usr/bin/env python3
-import tempfile
-import sys
-
+import os
+import pathlib
 from modeling_ddpm import DDPM
-
-model_id = sys.argv[1]
-
-ddpm = DDPM.from_pretrained(model_id)
-image = ddpm()
-
 import PIL.Image
 import numpy as np
-image_processed = image.cpu().permute(0, 2, 3, 1)
-image_processed = (image_processed + 1.0) * 127.5
-image_processed = image_processed.numpy().astype(np.uint8)
-image_pil = PIL.Image.fromarray(image_processed[0])
-image_pil.save("test.png")
 
-import ipdb; ipdb.set_trace()
+model_ids = ["ddpm-lsun-cat", "ddpm-lsun-cat-ema", "ddpm-lsun-church-ema", "ddpm-lsun-church", "ddpm-lsun-bedroom", "ddpm-lsun-bedroom-ema", "ddpm-cifar10-ema", "ddpm-cifar10", "ddpm-celeba-hq", "ddpm-celeba-hq-ema"]
+
+for model_id in model_ids:
+    path = os.path.join("/home/patrick/images/hf", model_id)
+    pathlib.Path(path).mkdir(parents=True, exist_ok=True)
+
+    ddpm = DDPM.from_pretrained("fusing/" + model_id)
+    image = ddpm(batch_size=4)
+
+    image_processed = image.cpu().permute(0, 2, 3, 1)
+    image_processed = (image_processed + 1.0) * 127.5
+    image_processed = image_processed.numpy().astype(np.uint8)
+
+    for i in range(image_processed.shape[0]):
+        image_pil = PIL.Image.fromarray(image_processed[i])
+        image_pil.save(os.path.join(path, f"image_{i}.png"))
diff --git a/models/vision/ddpm/modeling_ddpm.py b/models/vision/ddpm/modeling_ddpm.py
index ae049a8c0a..a10feaba40 100644
--- a/models/vision/ddpm/modeling_ddpm.py
+++ b/models/vision/ddpm/modeling_ddpm.py
@@ -27,12 +27,13 @@ class DDPM(DiffusionPipeline):
         super().__init__()
         self.register_modules(unet=unet, noise_scheduler=noise_scheduler)
 
-    def __call__(self, generator=None, torch_device=None):
-        torch_device = "cuda" if torch.cuda.is_available() else "cpu"
+    def __call__(self, batch_size=1, generator=None, torch_device=None):
+        if torch_device is None:
+            torch_device = "cuda" if torch.cuda.is_available() else "cpu"
 
         self.unet.to(torch_device)
         # 1. Sample gaussian noise
-        image = self.noise_scheduler.sample_noise((1, self.unet.in_channels, self.unet.resolution, self.unet.resolution), device=torch_device, generator=generator)
+        image = self.noise_scheduler.sample_noise((batch_size, self.unet.in_channels, self.unet.resolution, self.unet.resolution), device=torch_device, generator=generator)
         for t in tqdm.tqdm(reversed(range(len(self.noise_scheduler))), total=len(self.noise_scheduler)):
             # i) define coefficients for time step t
             clip_image_coeff = 1 / torch.sqrt(self.noise_scheduler.get_alpha_prod(t))
diff --git a/src/diffusers/schedulers/gaussian_ddpm.py b/src/diffusers/schedulers/gaussian_ddpm.py
index 4fcdfdf2bd..2a25cbbfc9 100644
--- a/src/diffusers/schedulers/gaussian_ddpm.py
+++ b/src/diffusers/schedulers/gaussian_ddpm.py
@@ -108,7 +108,7 @@ class GaussianDDPMScheduler(nn.Module, ConfigMixin):
 
     def sample_variance(self, time_step, shape, device, generator=None):
         variance = self.log_variance[time_step]
-        nonzero_mask = torch.tensor([1 - (time_step == 0)], device=device).float()[None, :].repeat(shape[0], 1)
+        nonzero_mask = torch.tensor([1 - (time_step == 0)], device=device).float()[None, :]
 
         noise = self.sample_noise(shape, device=device, generator=generator)
 
diff --git a/tests/test_modeling_utils.py b/tests/test_modeling_utils.py
index 4655c96749..6dce91ae4b 100755
--- a/tests/test_modeling_utils.py
+++ b/tests/test_modeling_utils.py
@@ -76,7 +76,7 @@ def floats_tensor(shape, scale=1.0, rng=None, name=None):
 class ModelTesterMixin(unittest.TestCase):
     @property
     def dummy_input(self):
-        batch_size = 1
+        batch_size = 4
         num_channels = 3
         sizes = (32, 32)