Skip to main content

> ML_ALGORITHM // CONSISTENCY-MODELS_v1.0

Consistency Models

Generative paradigm that maps arbitrary points at any noise level directly to the clean trajectory origin, enabling single-step generation.

Diffusion & Score-Based Generative Modelsdeep-generativeblack-boxlarge (>100k)
Back to All Algorithms
Computational Complexity
Training Complexity:O(epochs * batch_size * model_pass)
Inference Complexity:O(1 to 2 * model_pass)
Hardware Profile
CPU Friendly:No
Requires GPU:Yes
Memory Footprint:high
Interpretability & Data
Interpretability Tier:black-box
Training Data Needs:large (>100k)

Interpretability Assessment

Achieves high image synthesis quality in 1 or 2 forward passes without multi-step numerical integration.

Suitable Tasks & Supported Modalities

Suitable Tasks:
fast image generationsingle step sampling
Supported Modalities:
image

Implementing Libraries

diffusers
PyTorchLinux Foundation / PyTorch Foundation · v2.4.1
View Spec

Foundational Literature

Common Pitfalls & Warnings
  • Consistency training from scratch is significantly harder and lower quality than consistency distillation from a pretrained diffusion teacher