Noise is an image-shaped starting point filled with random values rather than meaningful visual content. Diffusion generation progressively transforms it into a high-signal, low-noise image. section 17.2.5 section 17.4
Starting a generation run
A generator creates a random tensor shaped like one or more RGB images, then repeatedly calls denoise(). With text conditioning, the same process can turn pure noise into an image matching a previously unseen prompt. section 17.2.5 section 17.3
def generate(self, num_images, diffusion_steps): noisy_images = keras.random.normal( (num_images, self.image_size, self.image_size, 3), #1 seed=self.seed_generator, #1 ) #1 step_size = 1.0 / diffusion_steps for step in range(diffusion_steps): diffusion_times = ops.ones((num_images, 1, 1, 1)) - step * step_size#2 noise_rates, signal_rates = diffusion_schedule(diffusion_times) #2 pred_images, pred_noises = self.denoise( #3 noisy_images, noise_rates, signal_rates #3 ) #3 next_diffusion_times = diffusion_times - step_size #4 next_noise_rates, next_signal_rates = diffusion_schedule( #4 next_diffusion_times #4 ) #4 noisy_images = ( next_signal_rates * pred_images + next_noise_rates * pred_noises ) images = ( #5 self.normalizer.mean + pred_images * self.normalizer.variance**0.5 #5 ) #5 return ops.clip(images, 0.0, 255.0)
Inspect the generation loop that initializes random image values and updates them through successive denoising steps. section 17.2.5
Use noise as the practical starting point for diffusion generation:
- Initialize a random image-shaped tensor. section 17.2.5
- Apply denoising repeatedly rather than expecting a finished image in one pass. section 17.2.5
- Condition denoising with text embeddings when generating text-to-image outputs. section 17.3
