Consistency model
Model Families AIMech:ConsistencyModel DRAFT
Aliases: Consistency models, CM (consistency model)
A generative model family that learns a noise- or time-conditioned function mapping points on a probability-flow ODE trajectory to its common low-noise endpoint, enabling a sample to be obtained with one model evaluation and allowing optional multistep refinement.
Identity Notes & Concept Scope
All 442 pinned AIO labels and scoped synonyms and 90 preceding native identities were checked, with ignored-inclusive text searches of both checkouts. The identity is the endpoint-map model family introduced by Song and colleagues, not generic logical consistency, a trained checkpoint, or one particular training loss. Original consistency distillation, original consistency training and improved consistency training are procedures for learning this family and are not aliases for the model family itself.
Claims & Grounded Evidence
2 Claims
Mechanism
#generative-mechanism
A consistency function assigns the same low-noise endpoint to states on the same probability-flow trajectory. The paper's skip-connected parameterization combines an input skip term and a learned term whose coefficients enforce the identity boundary at the minimum noise level. Generation applies the model to high-noise input; multistep generation interleaves model evaluation with noise injection. The original paper describes both diffusion-model-assisted distillation and direct consistency training with shared-noise adjacent perturbations of a data sample.
Limitation
#original-training-theory-limitation
The later paper gives a point-mass-data counterexample in which the limiting consistency-training/matching objective with unequal teacher and student parameters loses dependence on the data location. It motivates zero teacher EMA decay, with a stop-gradient teacher sharing current student parameters, for consistency training; this is distinct from averaging the model used for inference.
⚠️ Curated Limitations & Negative Results
- The learned consistency model is not required to be invertible. Ideal probability-flow properties do not establish exact data recovery or perfect self-consistency for a finite trained model.
- The original direct consistency-training teacher uses EMA; the later inspected correction identifies a failure of its original asymptotic justification when teacher and student parameters differ. This record does not endorse that original training argument as generally valid.
- The correction concerns consistency training and does not remove EMA from all distillation procedures or from the separate inference-time averaged model.
- One-evaluation generation is an available procedure, not a universal image-quality guarantee. No benchmark reproduction or independently verified convergence proof is claimed.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:Model
Model
|
broad_match | The consistency-model family is a specific learned representation of a generative distribution/endpoint transformation, within the broader scope of AIO Model. The mapping does not assert equivalence to a distillation procedure or to an adversarial network. |
Evidence Sources & Primary Literature
2 Sources| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
original-paper
|
Yang Song, Prafulla Dhariwal, Mark Chen, Ilya Sutskever
|
primary paper
full_text
|
ID: https://proceedings.mlr.press/v202/song23a.html
sha256:8f65127c770a…
|
training-correction
|
Yang Song, Prafulla Dhariwal
|
primary paper
full_text
|
ID: arXiv:2310.14189
sha256:2c77b5ae689b…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T04:42:32Z
Updated: 2026-10-11T04:42:32Z
Source File:
data/curated/ConsistencyModel.yaml