Implementation:Zai org CogVideo CogVideoXDPMScheduler From Config

Overview

Concrete tool for configuring the CogVideoX DPM noise scheduler for inference provided by the diffusers library. The scheduler controls the denoising step function and timestep selection during video generation.

Source

inference/cli_demo.py:L140-142

Signature

pipe.scheduler = CogVideoXDPMScheduler.from_config(
    pipe.scheduler.config,
    timestep_spacing: str = "trailing"
) -> CogVideoXDPMScheduler

Alternative for 2B Models

pipe.scheduler = CogVideoXDDIMScheduler.from_config(
    pipe.scheduler.config,
    timestep_spacing: str = "trailing"
) -> CogVideoXDDIMScheduler

Key Parameters

Parameter	Type	Default	Description
config	dict	(required)	Scheduler configuration from `pipe.scheduler.config`. Carries over all training-time scheduler settings.
timestep_spacing	str	"trailing"	Timestep spacing strategy. Use `"trailing"` for proper alignment with CogVideoX training schedule.

Inputs

Existing scheduler config -- The configuration dictionary from the pipeline's default scheduler (pipe.scheduler.config), which contains training-time parameters such as beta schedule, prediction type, and number of training timesteps.

Outputs

Configured scheduler instance -- A CogVideoXDPMScheduler (or CogVideoXDDIMScheduler) instance that is assigned directly to pipe.scheduler.

Usage Example

from diffusers import CogVideoXPipeline, CogVideoXDPMScheduler, CogVideoXDDIMScheduler
import torch

# Load pipeline
pipe = CogVideoXPipeline.from_pretrained(
    "THUDM/CogVideoX-5b",
    torch_dtype=torch.bfloat16
)

# For 5B models: use DPM scheduler
pipe.scheduler = CogVideoXDPMScheduler.from_config(
    pipe.scheduler.config,
    timestep_spacing="trailing"
)

# For 2B models (alternative): use DDIM scheduler
# pipe.scheduler = CogVideoXDDIMScheduler.from_config(
#     pipe.scheduler.config,
#     timestep_spacing="trailing"
# )

Import

from diffusers import CogVideoXDPMScheduler, CogVideoXDDIMScheduler

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment