Spaces:
Sleeping
Sleeping
File size: 6,864 Bytes
cf1a3cd 3338ab8 cf1a3cd | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 | """
Interpolation Module
Handles latent space interpolation for smooth frame transitions.
"""
import torch
from diffusers import AutoencoderKL, StableDiffusionPipeline
from PIL import Image
from typing import List, Optional
import numpy as np
class LatentInterpolator:
"""
Interpolates between keyframes in latent space for smooth animations.
"""
def __init__(
self,
model_id: str = "Manojb/stable-diffusion-2-1-base",
device: str = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu",
torch_dtype: torch.dtype = torch.float16 if torch.cuda.is_available() else torch.float32
):
"""
Initialize the interpolator.
Args:
model_id: Stable Diffusion model for VAE
device: Device to run on
torch_dtype: Data type for tensors
"""
self.device = device
self.torch_dtype = torch_dtype
print(f"Loading VAE from: {model_id}")
# Load VAE for encoding/decoding
self.vae = AutoencoderKL.from_pretrained(
model_id,
subfolder="vae",
torch_dtype=torch_dtype
).to(device)
print("VAE loaded successfully!")
def encode_image(self, image: Image.Image) -> torch.Tensor:
"""
Encode PIL Image to latent representation.
Args:
image: PIL Image to encode
Returns:
Latent tensor
"""
# Preprocess image
image = image.resize((512, 512))
image_np = np.array(image).astype(np.float32) / 255.0
image_np = image_np * 2.0 - 1.0 # Normalize to [-1, 1]
# Convert to tensor
image_tensor = torch.from_numpy(image_np).permute(2, 0, 1).unsqueeze(0)
image_tensor = image_tensor.to(device=self.device, dtype=self.torch_dtype)
# Encode to latent space
with torch.no_grad():
latent = self.vae.encode(image_tensor).latent_dist.sample()
latent = latent * 0.18215 # Scaling factor for SD
return latent
def decode_latent(self, latent: torch.Tensor) -> Image.Image:
"""
Decode latent representation to PIL Image.
Args:
latent: Latent tensor
Returns:
PIL Image
"""
# Decode from latent space
with torch.no_grad():
latent = latent / 0.18215
image_tensor = self.vae.decode(latent).sample
# Post-process
image_tensor = (image_tensor / 2 + 0.5).clamp(0, 1)
image_np = image_tensor.cpu().permute(0, 2, 3, 1).numpy()[0]
image_np = (image_np * 255).astype(np.uint8)
return Image.fromarray(image_np)
def interpolate_latents(
self,
latent_start: torch.Tensor,
latent_end: torch.Tensor,
num_steps: int = 10,
interpolation_type: str = "linear"
) -> List[torch.Tensor]:
"""
Interpolate between two latent representations.
Args:
latent_start: Starting latent
latent_end: Ending latent
num_steps: Number of interpolation steps
interpolation_type: Type of interpolation (linear, slerp)
Returns:
List of interpolated latents
"""
interpolated = []
for i in range(num_steps):
alpha = i / (num_steps - 1) if num_steps > 1 else 0
if interpolation_type == "linear":
# Linear interpolation
latent = (1 - alpha) * latent_start + alpha * latent_end
elif interpolation_type == "slerp":
# Spherical linear interpolation (better for latent spaces)
latent = self._slerp(latent_start, latent_end, alpha)
else:
raise ValueError(f"Unknown interpolation type: {interpolation_type}")
interpolated.append(latent)
return interpolated
def _slerp(
self,
v0: torch.Tensor,
v1: torch.Tensor,
t: float,
eps: float = 1e-8
) -> torch.Tensor:
"""
Spherical linear interpolation.
Args:
v0: Start vector
v1: End vector
t: Interpolation parameter [0, 1]
eps: Small value to prevent division by zero
Returns:
Interpolated tensor
"""
# Normalize vectors
v0_norm = v0 / (torch.norm(v0) + eps)
v1_norm = v1 / (torch.norm(v1) + eps)
# Compute angle
dot = (v0_norm * v1_norm).sum()
dot = torch.clamp(dot, -1.0, 1.0)
omega = torch.acos(dot)
# Compute interpolation
so = torch.sin(omega)
if so < eps:
# Vectors are nearly parallel, use linear interpolation
return (1.0 - t) * v0 + t * v1
return (torch.sin((1.0 - t) * omega) / so) * v0 + (torch.sin(t * omega) / so) * v1
def generate_interpolated_sequence(
self,
keyframes: List[Image.Image],
frames_between: int = 4,
interpolation_type: str = "slerp"
) -> List[Image.Image]:
"""
Generate smooth sequence by interpolating between keyframes.
Args:
keyframes: List of keyframe images
frames_between: Number of frames to generate between each keyframe
interpolation_type: Interpolation method
Returns:
List of all frames (keyframes + interpolated)
"""
if len(keyframes) < 2:
return keyframes
all_frames = []
for i in range(len(keyframes) - 1):
# Encode keyframes to latent space
latent_start = self.encode_image(keyframes[i])
latent_end = self.encode_image(keyframes[i + 1])
# Interpolate
interpolated_latents = self.interpolate_latents(
latent_start,
latent_end,
num_steps=frames_between + 2, # +2 to include start and end
interpolation_type=interpolation_type
)
# Decode all except the last (avoid duplicates)
for latent in interpolated_latents[:-1]:
frame = self.decode_latent(latent)
all_frames.append(frame)
# Add final keyframe
all_frames.append(keyframes[-1])
return all_frames
if __name__ == "__main__":
print("Latent interpolation module loaded.")
print("Use with keyframe images to generate smooth interpolations.")
|