Skip to content

Linear Noise Schedule

Implementation of a linear noise schedule for diffusion models.

LinearNoiseSchedule

Bases: BaseNoiseSchedule

Linear noise schedule that increases linearly over time.

This class implements a simple linear noise schedule where the noise increases linearly from beta_min to beta_max over the diffusion process.

Source code in image_gen\noise\linear.py 
 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
class LinearNoiseSchedule(BaseNoiseSchedule):
    """Linear noise schedule that increases linearly over time.

    This class implements a simple linear noise schedule where the noise
    increases linearly from beta_min to beta_max over the diffusion process.
    """

    def __init__(
        self,
        *_,
        beta_min: float = 0.0001,
        beta_max: float = 20.0,
        **__
    ):
        """Initialize the linear noise schedule.

        Args:
            beta_min: Minimum noise value at t=0. Defaults to 0.0001.
            beta_max: Maximum noise value at t=1. Defaults to 20.0.
        """
        self.beta_min = beta_min
        self.beta_max = beta_max

    def __call__(self, t: Tensor, *_, **__) -> Tensor:
        """Calculate noise at specific timesteps.

        Args:
            t: Tensor containing timestep values in range [0, 1].

        Returns:
            Tensor: Noise values corresponding to the input timesteps.
        """
        return self.beta_min + t * (self.beta_max - self.beta_min)

    def integral_beta(self, t: Tensor, *_, **__) -> Tensor:
        """Calculate the integral of the noise function up to timestep t.

        The analytical solution for the integral of a linear function
        from 0 to t is: beta_min * t + 0.5 * (beta_max - beta_min) * t^2.

        Args:
            t: Tensor containing timestep values in range [0, 1].

        Returns:
            Tensor: Integrated noise values corresponding to the input timesteps.
        """
        return self.beta_min * t + 0.5 * (self.beta_max - self.beta_min) * (t ** 2)

    def config(self) -> dict:
        """Get the configuration parameters of the noise schedule.

        Returns:
            dict: Dictionary containing the configuration parameters.
        """
        return {
            "beta_min": self.beta_min,
            "beta_max": self.beta_max,
        }

__call__(t, *_, **__)

Calculate noise at specific timesteps.

Parameters:

Name Type Description Default
t Tensor

Tensor containing timestep values in range [0, 1].

 required

Returns:

Name Type Description
Tensor Tensor

Noise values corresponding to the input timesteps.
Source code in image_gen\noise\linear.py 
31
32
33
34
35
36
37
38
39
40
def __call__(self, t: Tensor, *_, **__) -> Tensor:
    """Calculate noise at specific timesteps.

    Args:
        t: Tensor containing timestep values in range [0, 1].

    Returns:
        Tensor: Noise values corresponding to the input timesteps.
    """
    return self.beta_min + t * (self.beta_max - self.beta_min)

__init__(*_, beta_min=0.0001, beta_max=20.0, **__)

Initialize the linear noise schedule.

Parameters:

Name Type Description Default
beta_min float

Minimum noise value at t=0. Defaults to 0.0001.

 0.0001
beta_max float

Maximum noise value at t=1. Defaults to 20.0.

 20.0
Source code in image_gen\noise\linear.py 
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
def __init__(
    self,
    *_,
    beta_min: float = 0.0001,
    beta_max: float = 20.0,
    **__
):
    """Initialize the linear noise schedule.

    Args:
        beta_min: Minimum noise value at t=0. Defaults to 0.0001.
        beta_max: Maximum noise value at t=1. Defaults to 20.0.
    """
    self.beta_min = beta_min
    self.beta_max = beta_max

config()

Get the configuration parameters of the noise schedule.

Returns:

Name Type Description
dict dict

Dictionary containing the configuration parameters.
Source code in image_gen\noise\linear.py 
56
57
58
59
60
61
62
63
64
65
def config(self) -> dict:
    """Get the configuration parameters of the noise schedule.

    Returns:
        dict: Dictionary containing the configuration parameters.
    """
    return {
        "beta_min": self.beta_min,
        "beta_max": self.beta_max,
    }

integral_beta(t, *_, **__)

Calculate the integral of the noise function up to timestep t.

The analytical solution for the integral of a linear function from 0 to t is: beta_min * t + 0.5 * (beta_max - beta_min) * t^2.

Parameters:

Name Type Description Default
t Tensor

Tensor containing timestep values in range [0, 1].

 required

Returns:

Name Type Description
Tensor Tensor

Integrated noise values corresponding to the input timesteps.
Source code in image_gen\noise\linear.py 
42
43
44
45
46
47
48
49
50
51
52
53
54
def integral_beta(self, t: Tensor, *_, **__) -> Tensor:
    """Calculate the integral of the noise function up to timestep t.

    The analytical solution for the integral of a linear function
    from 0 to t is: beta_min * t + 0.5 * (beta_max - beta_min) * t^2.

    Args:
        t: Tensor containing timestep values in range [0, 1].

    Returns:
        Tensor: Integrated noise values corresponding to the input timesteps.
    """
    return self.beta_min * t + 0.5 * (self.beta_max - self.beta_min) * (t ** 2)