Denoising Diffusion Implicit Models (DDIM)
Overview
Denoising Diffusion Implicit Models (DDIM) by Jiaming Song, Chenlin Meng and Stefano Ermon.
The abstract of the paper is the following:
Denoising diffusion probabilistic models (DDPMs) have achieved high quality image generation without adversarial training, yet they require simulating a Markov chain for many steps to produce a sample. To accelerate sampling, we present denoising diffusion implicit models (DDIMs), a more efficient class of iterative implicit probabilistic models with the same training procedure as DDPMs. In DDPMs, the generative process is defined as the reverse of a Markovian diffusion process. We construct a class of non-Markovian diffusion processes that lead to the same training objective, but whose reverse process can be much faster to sample from. We empirically demonstrate that DDIMs can produce high quality samples 10× to 50× faster in terms of wall-clock time compared to DDPMs, allow us to trade off computation for sample quality, and can perform semantically meaningful image interpolation directly in the latent space.
The original codebase of this paper can be found here: ermongroup/ddim. For questions, feel free to contact the author on tsong.me.
DDIMScheduler
class diffusers.DDIMScheduler
< source >( num_train_timesteps: int = 1000 beta_start: float = 0.0001 beta_end: float = 0.02 beta_schedule: str = 'linear' trained_betas: typing.Union[numpy.ndarray, typing.List[float], NoneType] = None clip_sample: bool = True set_alpha_to_one: bool = True steps_offset: int = 0 prediction_type: str = 'epsilon' thresholding: bool = False dynamic_thresholding_ratio: float = 0.995 clip_sample_range: float = 1.0 sample_max_value: float = 1.0 )
Parameters
-
num_train_timesteps (
int
) — number of diffusion steps used to train the model. -
beta_start (
float
) — the startingbeta
value of inference. -
beta_end (
float
) — the finalbeta
value. -
beta_schedule (
str
) — the beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose fromlinear
,scaled_linear
, orsquaredcos_cap_v2
. -
trained_betas (
np.ndarray
, optional) — option to pass an array of betas directly to the constructor to bypassbeta_start
,beta_end
etc. -
clip_sample (
bool
, defaultTrue
) — option to clip predicted sample for numerical stability. -
clip_sample_range (
float
, default1.0
) — the maximum magnitude for sample clipping. Valid only whenclip_sample=True
. -
set_alpha_to_one (
bool
, defaultTrue
) — each diffusion step uses the value of alphas product at that step and at the previous one. For the final step there is no previous alpha. When this option isTrue
the previous alpha product is fixed to1
, otherwise it uses the value of alpha at step 0. -
steps_offset (
int
, default0
) — an offset added to the inference steps. You can use a combination ofoffset=1
andset_alpha_to_one=False
, to make the last step use step 0 for the previous alpha product, as done in stable diffusion. -
prediction_type (
str
, defaultepsilon
, optional) — prediction type of the scheduler function, one ofepsilon
(predicting the noise of the diffusion process),sample
(directly predicting the noisy sample) or
v_prediction` (see section 2.4 https://imagen.research.google/video/paper.pdf) -
thresholding (
bool
, defaultFalse
) — whether to use the “dynamic thresholding” method (introduced by Imagen, https://arxiv.org/abs/2205.11487). Note that the thresholding method is unsuitable for latent-space diffusion models (such as stable-diffusion). -
dynamic_thresholding_ratio (
float
, default0.995
) — the ratio for the dynamic thresholding method. Default is0.995
, the same as Imagen (https://arxiv.org/abs/2205.11487). Valid only whenthresholding=True
. -
sample_max_value (
float
, default1.0
) — the threshold value for dynamic thresholding. Valid only whenthresholding=True
.
Denoising diffusion implicit models is a scheduler that extends the denoising procedure introduced in denoising diffusion probabilistic models (DDPMs) with non-Markovian guidance.
~ConfigMixin takes care of storing all config attributes that are passed in the scheduler’s __init__
function, such as num_train_timesteps
. They can be accessed via scheduler.config.num_train_timesteps
.
SchedulerMixin provides general loading and saving functionality via the SchedulerMixin.save_pretrained() and
from_pretrained() functions.
For more details, see the original paper: https://arxiv.org/abs/2010.02502
scale_model_input
< source >(
sample: FloatTensor
timestep: typing.Optional[int] = None
)
→
torch.FloatTensor
Ensures interchangeability with schedulers that need to scale the denoising model input depending on the current timestep.
set_timesteps
< source >( num_inference_steps: int device: typing.Union[str, torch.device] = None )
Sets the discrete timesteps used for the diffusion chain. Supporting function to be run before inference.
step
< source >(
model_output: FloatTensor
timestep: int
sample: FloatTensor
eta: float = 0.0
use_clipped_model_output: bool = False
generator = None
variance_noise: typing.Optional[torch.FloatTensor] = None
return_dict: bool = True
)
→
~schedulers.scheduling_utils.DDIMSchedulerOutput
or tuple
Parameters
-
model_output (
torch.FloatTensor
) — direct output from learned diffusion model. -
timestep (
int
) — current discrete timestep in the diffusion chain. -
sample (
torch.FloatTensor
) — current instance of sample being created by diffusion process. -
eta (
float
) — weight of noise for added noise in diffusion step. -
use_clipped_model_output (
bool
) — ifTrue
, compute “corrected”model_output
from the clipped predicted original sample. Necessary because predicted original sample is clipped to [-1, 1] whenself.config.clip_sample
isTrue
. If no clipping has happened, “corrected”model_output
would coincide with the one provided as input anduse_clipped_model_output
will have not effect. generator — random number generator. -
variance_noise (
torch.FloatTensor
) — instead of generating noise for the variance usinggenerator
, we can directly provide the noise for the variance itself. This is useful for methods such as CycleDiffusion. (https://arxiv.org/abs/2210.05559) -
return_dict (
bool
) — option for returning tuple rather than DDIMSchedulerOutput class
Returns
~schedulers.scheduling_utils.DDIMSchedulerOutput
or tuple
~schedulers.scheduling_utils.DDIMSchedulerOutput
if return_dict
is True, otherwise a tuple
. When
returning a tuple, the first element is the sample tensor.
Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion process from the learned model outputs (most often the predicted noise).