|
| 1 | +# SPDX-License-Identifier: Apache-2.0 |
| 2 | +""" |
| 3 | +Minimal image processing utilities for FastVideo. |
| 4 | +This module provides lightweight image preprocessing without external dependencies beyond PyTorch/NumPy/PIL. |
| 5 | +""" |
| 6 | + |
| 7 | +from typing import Optional, Union |
| 8 | + |
| 9 | +import numpy as np |
| 10 | +import PIL.Image |
| 11 | +import torch |
| 12 | + |
| 13 | + |
| 14 | +class ImageProcessor: |
| 15 | + """ |
| 16 | + Minimal image processor for video frame preprocessing. |
| 17 | +
|
| 18 | + This is a lightweight alternative to diffusers.VideoProcessor that handles: |
| 19 | + - PIL image to tensor conversion |
| 20 | + - Resizing to specified dimensions |
| 21 | + - Normalization to [-1, 1] range |
| 22 | +
|
| 23 | + Args: |
| 24 | + vae_scale_factor: The VAE scale factor used to ensure dimensions are multiples of this value. |
| 25 | + """ |
| 26 | + |
| 27 | + def __init__(self, vae_scale_factor: int = 8) -> None: |
| 28 | + self.vae_scale_factor = vae_scale_factor |
| 29 | + |
| 30 | + def preprocess( |
| 31 | + self, |
| 32 | + image: Union[PIL.Image.Image, np.ndarray, torch.Tensor], |
| 33 | + height: Optional[int] = None, |
| 34 | + width: Optional[int] = None, |
| 35 | + ) -> torch.Tensor: |
| 36 | + """ |
| 37 | + Preprocess an image to a normalized torch tensor. |
| 38 | +
|
| 39 | + Args: |
| 40 | + image: Input image (PIL Image, NumPy array, or torch tensor) |
| 41 | + height: Target height. If None, uses image's original height. |
| 42 | + width: Target width. If None, uses image's original width. |
| 43 | +
|
| 44 | + Returns: |
| 45 | + torch.Tensor: Normalized tensor of shape (1, 3, height, width) or (1, 1, height, width) for grayscale, |
| 46 | + with values in range [-1, 1]. |
| 47 | + """ |
| 48 | + # Handle different input types |
| 49 | + if isinstance(image, PIL.Image.Image): |
| 50 | + return self._preprocess_pil(image, height, width) |
| 51 | + elif isinstance(image, np.ndarray): |
| 52 | + return self._preprocess_numpy(image, height, width) |
| 53 | + elif isinstance(image, torch.Tensor): |
| 54 | + return self._preprocess_tensor(image, height, width) |
| 55 | + else: |
| 56 | + raise ValueError( |
| 57 | + f"Unsupported image type: {type(image)}. " |
| 58 | + "Supported types: PIL.Image.Image, np.ndarray, torch.Tensor" |
| 59 | + ) |
| 60 | + |
| 61 | + def _preprocess_pil( |
| 62 | + self, |
| 63 | + image: PIL.Image.Image, |
| 64 | + height: Optional[int] = None, |
| 65 | + width: Optional[int] = None, |
| 66 | + ) -> torch.Tensor: |
| 67 | + """Preprocess a PIL image.""" |
| 68 | + if height is None: |
| 69 | + height = image.height |
| 70 | + if width is None: |
| 71 | + width = image.width |
| 72 | + |
| 73 | + height = height - (height % self.vae_scale_factor) |
| 74 | + width = width - (width % self.vae_scale_factor) |
| 75 | + |
| 76 | + image = image.resize((width, height), resample=PIL.Image.Resampling.LANCZOS) |
| 77 | + |
| 78 | + image_np = np.array(image, dtype=np.float32) / 255.0 |
| 79 | + |
| 80 | + if image_np.ndim == 2: # Grayscale |
| 81 | + image_np = np.expand_dims(image_np, axis=-1) |
| 82 | + |
| 83 | + return self._normalize_to_tensor(image_np) |
| 84 | + |
| 85 | + def _preprocess_numpy( |
| 86 | + self, |
| 87 | + image: np.ndarray, |
| 88 | + height: Optional[int] = None, |
| 89 | + width: Optional[int] = None, |
| 90 | + ) -> torch.Tensor: |
| 91 | + """Preprocess a numpy array.""" |
| 92 | + # Determine target dimensions if not provided |
| 93 | + if image.ndim == 3: |
| 94 | + img_height, img_width = image.shape[:2] |
| 95 | + elif image.ndim == 2: |
| 96 | + img_height, img_width = image.shape |
| 97 | + else: |
| 98 | + raise ValueError(f"Expected 2D or 3D array, got {image.ndim}D") |
| 99 | + |
| 100 | + if height is None: |
| 101 | + height = img_height |
| 102 | + if width is None: |
| 103 | + width = img_width |
| 104 | + |
| 105 | + height = height - (height % self.vae_scale_factor) |
| 106 | + width = width - (width % self.vae_scale_factor) |
| 107 | + |
| 108 | + if image.dtype == np.uint8: |
| 109 | + pil_image = PIL.Image.fromarray(image) |
| 110 | + else: |
| 111 | + # Assume normalized [0, 1] or similar |
| 112 | + if image.max() <= 1.0: |
| 113 | + image_uint8 = (image * 255).astype(np.uint8) |
| 114 | + else: |
| 115 | + image_uint8 = image.astype(np.uint8) |
| 116 | + pil_image = PIL.Image.fromarray(image_uint8) |
| 117 | + |
| 118 | + pil_image = pil_image.resize((width, height), resample=PIL.Image.Resampling.LANCZOS) |
| 119 | + image_np = np.array(pil_image, dtype=np.float32) / 255.0 |
| 120 | + |
| 121 | + # Ensure 3D shape |
| 122 | + if image_np.ndim == 2: |
| 123 | + image_np = np.expand_dims(image_np, axis=-1) |
| 124 | + |
| 125 | + return self._normalize_to_tensor(image_np) |
| 126 | + |
| 127 | + def _preprocess_tensor( |
| 128 | + self, |
| 129 | + image: torch.Tensor, |
| 130 | + height: Optional[int] = None, |
| 131 | + width: Optional[int] = None, |
| 132 | + ) -> torch.Tensor: |
| 133 | + """Preprocess a torch tensor.""" |
| 134 | + # Determine target dimensions |
| 135 | + if image.ndim == 3: # (H, W, C) or (C, H, W) |
| 136 | + if image.shape[0] in (1, 3, 4): # Likely (C, H, W) |
| 137 | + img_height, img_width = image.shape[1], image.shape[2] |
| 138 | + else: # Likely (H, W, C) |
| 139 | + img_height, img_width = image.shape[0], image.shape[1] |
| 140 | + elif image.ndim == 2: # (H, W) |
| 141 | + img_height, img_width = image.shape |
| 142 | + else: |
| 143 | + raise ValueError(f"Expected 2D or 3D tensor, got {image.ndim}D") |
| 144 | + |
| 145 | + if height is None: |
| 146 | + height = img_height |
| 147 | + if width is None: |
| 148 | + width = img_width |
| 149 | + |
| 150 | + height = height - (height % self.vae_scale_factor) |
| 151 | + width = width - (width % self.vae_scale_factor) |
| 152 | + |
| 153 | + if image.ndim == 2: |
| 154 | + image = image.unsqueeze(0).unsqueeze(0) # (1, 1, H, W) |
| 155 | + elif image.ndim == 3: |
| 156 | + if image.shape[0] in (1, 3, 4): # (C, H, W) |
| 157 | + image = image.unsqueeze(0) # (1, C, H, W) |
| 158 | + else: # (H, W, C) - need to rearrange |
| 159 | + image = image.permute(2, 0, 1).unsqueeze(0) # (1, C, H, W) |
| 160 | + |
| 161 | + image = torch.nn.functional.interpolate( |
| 162 | + image, size=(height, width), mode="bilinear", align_corners=False |
| 163 | + ) |
| 164 | + |
| 165 | + if image.max() > 1.0: # Assume [0, 255] range |
| 166 | + image = image / 255.0 |
| 167 | + |
| 168 | + image = 2.0 * image - 1.0 |
| 169 | + |
| 170 | + return image |
| 171 | + |
| 172 | + def _normalize_to_tensor(self, image_np: np.ndarray) -> torch.Tensor: |
| 173 | + """ |
| 174 | + Convert normalized numpy array [0, 1] to torch tensor [-1, 1]. |
| 175 | +
|
| 176 | + Args: |
| 177 | + image_np: NumPy array with shape (H, W) or (H, W, C) with values in [0, 1] |
| 178 | +
|
| 179 | + Returns: |
| 180 | + torch.Tensor: Shape (1, C, H, W) or (1, 1, H, W) with values in [-1, 1] |
| 181 | + """ |
| 182 | + # Convert to tensor |
| 183 | + if image_np.ndim == 2: # (H, W) - grayscale |
| 184 | + tensor = torch.from_numpy(image_np).unsqueeze(0).unsqueeze(0) # (1, 1, H, W) |
| 185 | + elif image_np.ndim == 3: # (H, W, C) |
| 186 | + tensor = torch.from_numpy(image_np).permute(2, 0, 1).unsqueeze(0) # (1, C, H, W) |
| 187 | + else: |
| 188 | + raise ValueError(f"Expected 2D or 3D array, got {image_np.ndim}D") |
| 189 | + |
| 190 | + # Normalize to [-1, 1] |
| 191 | + tensor = 2.0 * tensor - 1.0 |
| 192 | + |
| 193 | + return tensor |
0 commit comments