TCDScheduler
Trajectory Consistency Distillation by Jianbin Zheng, Minghui Hu, Zhongyi Fan, Chaoyue Wang, Changxing Ding, Dacheng Tao and Tat-Jen Cham introduced a Strategic Stochastic Sampling (Algorithm 4) that is capable of generating good samples in a small number of steps. Distinguishing it as an advanced iteration of the multistep scheduler (Algorithm 1) in the Consistency Models, Strategic Stochastic Sampling specifically tailored for the trajectory consistency function.
The abstract from the paper is:
Latent Consistency Model (LCM) extends the Consistency Model to the latent space and leverages the guided consistency distillation technique to achieve impressive performance in accelerating text-to-image synthesis. However, we observed that LCM struggles to generate images with both clarity and detailed intricacy. To address this limitation, we initially delve into and elucidate the underlying causes. Our investigation identifies that the primary issue stems from errors in three distinct areas. Consequently, we introduce Trajectory Consistency Distillation (TCD), which encompasses trajectory consistency function and strategic stochastic sampling. The trajectory consistency function diminishes the distillation errors by broadening the scope of the self-consistency boundary condition and endowing the TCD with the ability to accurately trace the entire trajectory of the Probability Flow ODE. Additionally, strategic stochastic sampling is specifically designed to circumvent the accumulated errors inherent in multi-step consistency sampling, which is meticulously tailored to complement the TCD model. Experiments demonstrate that TCD not only significantly enhances image quality at low NFEs but also yields more detailed results compared to the teacher model at high NFEs.
The original codebase can be found at jabir-zheng/TCD.
TCDScheduler
class diffusers.TCDScheduler
< source >( num_train_timesteps: int = 1000 beta_start: float = 0.00085 beta_end: float = 0.012 beta_schedule: str = 'scaled_linear' trained_betas: Union = None original_inference_steps: int = 50 clip_sample: bool = False clip_sample_range: float = 1.0 set_alpha_to_one: bool = True steps_offset: int = 0 prediction_type: str = 'epsilon' thresholding: bool = False dynamic_thresholding_ratio: float = 0.995 sample_max_value: float = 1.0 timestep_spacing: str = 'leading' timestep_scaling: float = 10.0 rescale_betas_zero_snr: bool = False )
Parameters
- num_train_timesteps (
int
, defaults to 1000) — The number of diffusion steps to train the model. - beta_start (
float
, defaults to 0.0001) — The startingbeta
value of inference. - beta_end (
float
, defaults to 0.02) — The finalbeta
value. - beta_schedule (
str
, defaults to"linear"
) — 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) — Pass an array of betas directly to the constructor to bypassbeta_start
andbeta_end
. - original_inference_steps (
int
, optional, defaults to 50) — The default number of inference steps used to generate a linearly-spaced timestep schedule, from which we will ultimately takenum_inference_steps
evenly spaced timesteps to form the final timestep schedule. - clip_sample (
bool
, defaults toTrue
) — Clip the predicted sample for numerical stability. - clip_sample_range (
float
, defaults to 1.0) — The maximum magnitude for sample clipping. Valid only whenclip_sample=True
. - set_alpha_to_one (
bool
, defaults toTrue
) — Each diffusion step uses the alphas product value 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 alpha value at step 0. - steps_offset (
int
, defaults to 0) — An offset added to the inference steps, as required by some model families. - prediction_type (
str
, defaults toepsilon
, optional) — Prediction type of the scheduler function; can beepsilon
(predicts the noise of the diffusion process),sample
(directly predicts the noisy sample) or
v_prediction` (see section 2.4 of Imagen Video paper). - thresholding (
bool
, defaults toFalse
) — Whether to use the “dynamic thresholding” method. This is unsuitable for latent-space diffusion models such as Stable Diffusion. - dynamic_thresholding_ratio (
float
, defaults to 0.995) — The ratio for the dynamic thresholding method. Valid only whenthresholding=True
. - sample_max_value (
float
, defaults to 1.0) — The threshold value for dynamic thresholding. Valid only whenthresholding=True
. - timestep_spacing (
str
, defaults to"leading"
) — The way the timesteps should be scaled. Refer to Table 2 of the Common Diffusion Noise Schedules and Sample Steps are Flawed for more information. - timestep_scaling (
float
, defaults to 10.0) — The factor the timesteps will be multiplied by when calculating the consistency model boundary conditionsc_skip
andc_out
. Increasing this will decrease the approximation error (although the approximation error at the default of10.0
is already pretty small). - rescale_betas_zero_snr (
bool
, defaults toFalse
) — Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and dark samples instead of limiting it to samples with medium brightness. Loosely related to--offset_noise
.
TCDScheduler
incorporates the Strategic Stochastic Sampling
introduced by the paper Trajectory Consistency Distillation
, extending the original Multistep Consistency Sampling to enable unrestricted trajectory traversal.
This code is based on the official repo of TCD(https://github.com/jabir-zheng/TCD).
This model inherits from SchedulerMixin and ConfigMixin. ~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.
scale_model_input
< source >( sample: Tensor timestep: Optional = None ) → torch.Tensor
Ensures interchangeability with schedulers that need to scale the denoising model input depending on the current timestep.
set_begin_index
< source >( begin_index: int = 0 )
Sets the begin index for the scheduler. This function should be run from pipeline before the inference.
set_timesteps
< source >( num_inference_steps: Optional = None device: Union = None original_inference_steps: Optional = None timesteps: Optional = None strength: float = 1.0 )
Parameters
- num_inference_steps (
int
, optional) — The number of diffusion steps used when generating samples with a pre-trained model. If used,timesteps
must beNone
. - device (
str
ortorch.device
, optional) — The device to which the timesteps should be moved to. IfNone
, the timesteps are not moved. - original_inference_steps (
int
, optional) — The original number of inference steps, which will be used to generate a linearly-spaced timestep schedule (which is different from the standarddiffusers
implementation). We will then takenum_inference_steps
timesteps from this schedule, evenly spaced in terms of indices, and use that as our final timestep schedule. If not set, this will default to theoriginal_inference_steps
attribute. - timesteps (
List[int]
, optional) — Custom timesteps used to support arbitrary spacing between timesteps. IfNone
, then the default timestep spacing strategy of equal spacing between timesteps on the training/distillation timestep schedule is used. Iftimesteps
is passed,num_inference_steps
must beNone
. - strength (
float
, optional, defaults to 1.0) — Used to determine the number of timesteps used for inference when using img2img, inpaint, etc.
Sets the discrete timesteps used for the diffusion chain (to be run before inference).
step
< source >( model_output: Tensor timestep: int sample: Tensor eta: float = 0.3 generator: Optional = None return_dict: bool = True ) → ~schedulers.scheduling_utils.TCDSchedulerOutput
or tuple
Parameters
- model_output (
torch.Tensor
) — The direct output from learned diffusion model. - timestep (
int
) — The current discrete timestep in the diffusion chain. - sample (
torch.Tensor
) — A current instance of a sample created by the diffusion process. - eta (
float
) — A stochastic parameter (referred to asgamma
in the paper) used to control the stochasticity in every step. When eta = 0, it represents deterministic sampling, whereas eta = 1 indicates full stochastic sampling. - generator (
torch.Generator
, optional) — A random number generator. - return_dict (
bool
, optional, defaults toTrue
) — Whether or not to return a TCDSchedulerOutput ortuple
.
Returns
~schedulers.scheduling_utils.TCDSchedulerOutput
or tuple
If return_dict is True
, TCDSchedulerOutput is returned, otherwise a
tuple is returned where the first element is the sample tensor.
Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion process from the learned model outputs (most often the predicted noise).
TCDSchedulerOutput
class diffusers.schedulers.scheduling_tcd.TCDSchedulerOutput
< source >( prev_sample: Tensor pred_noised_sample: Optional = None )
Parameters
- prev_sample (
torch.Tensor
of shape(batch_size, num_channels, height, width)
for images) — Computed sample(x_{t-1})
of previous timestep.prev_sample
should be used as next model input in the denoising loop. - pred_noised_sample (
torch.Tensor
of shape(batch_size, num_channels, height, width)
for images) — The predicted noised sample(x_{s})
based on the model output from the current timestep.
Output class for the scheduler’s step
function output.