Generative Consistency Models

• Generative Consistency Models (CMs) are generative models that enforce self-consistency along a probability flow ODE trajectory, mapping noisy states to a clean endpoint.
• CMs achieve one-step or few-step inference by learning direct trajectory-invariant projections, drastically reducing the reverse diffusion steps required in traditional models.
• Extensions of CMs incorporate phased, manifold, and control-based adaptations to improve performance across image generation, channel estimation, text-to-speech, motion synthesis, and other domains.

Generative Consistency Models (CMs) are a class of generative models that enforce self-consistency along the probability flow ordinary differential equation (PF ODE) trajectory of a diffusion process, learning a mapping from noisy states at different noise levels to a common clean endpoint. Their defining practical property is that they can perform generative inference in one step, or in a few refinement steps if desired, rather than requiring the tens to hundreds of reverse-time steps typical of diffusion and score-based models. Since their introduction, CMs have developed from diffusion-distilled image generators into a broader family that includes phased, trajectory-aware, flow-anchored, manifold-valued, and application-specific variants for inverse problems, reinforcement learning, text-to-speech, motion synthesis, and scientific posterior estimation (Song et al., 2023, Kim et al., 2023).

1. Probability-flow formulation and the consistency constraint

The mathematical core of a CM is a time-indexed function defined on the PF ODE associated with a diffusion process. In the original formulation, if $\mathbf{x}_t$ and $\mathbf{x}_{t'}$ lie on the same PF ODE trajectory, then the model enforces

$$f_\theta(\mathbf{x}_t,t)=f_\theta(\mathbf{x}_{t'},t'),$$

with a boundary condition at the small-noise endpoint, typically written as $$f_\theta(\mathbf{x}_\epsilon,\epsilon)=\mathbf{x}_\epsilon$$. Across the literature, the endpoint is written as $\mathbf{x}_\epsilon$, $\mathbf{x}_0$, or $\mathbf{x}_1$, but the common idea is identical: every point on the same trajectory should map to the same clean target (Song et al., 2023, Li et al., 26 Apr 2026).

This condition is naturally tied to the PF ODE of a diffusion model. In the general score-based setting, the PF ODE is

$$\frac{d\mathbf{x}_t}{dt}=\mu(\mathbf{x}_t,t)-\frac{1}{2}\sigma(t)^2\nabla \log p_t(\mathbf{x}_t).$$

In the EDM setting used in early CM work, $\mu(\mathbf{x},t)=0$ and $\sigma(t)=\sqrt{2t}$, so the forward noising process becomes $\mathbf{x}_{t'}$0 with $\mathbf{x}_{t'}$1, and the empirical PF ODE can be written as $\mathbf{x}_{t'}$2 when a score model $\mathbf{x}_{t'}$3 is available (Song et al., 2023).

A practical parameterization enforces the boundary condition through a skip connection:

$\mathbf{x}_{t'}$4

with $\mathbf{x}_{t'}$5 and $\mathbf{x}_{t'}$6. This makes the small-noise endpoint exact by construction while leaving $\mathbf{x}_{t'}$7 unconstrained away from the boundary. Variants of this parameterization recur across image, channel-estimation, PET, and TTS settings, often with domain-specific preconditioning constants or architecture choices (Li et al., 26 Apr 2026, Li et al., 2024).

The resulting distinction from diffusion is structural rather than merely algorithmic. Diffusion models learn local denoising or score fields and then numerically integrate a reverse trajectory; CMs learn a direct trajectory-invariant projector. This is why one-step generation is not an afterthought in CMs but part of the model definition itself (Song et al., 2023, Geng et al., 2024).

2. Training objectives and sampling regimes

Two training paradigms dominate the CM literature. Consistency distillation (CD) uses a pretrained diffusion teacher or a numerical PF ODE solver to generate neighboring trajectory points and trains the student to make their outputs agree. Consistency training (CT) removes the teacher and enforces self-consistency directly across adjacent noisy observations formed from the same clean sample and shared noise. A standard CT objective is

$\mathbf{x}_{t'}$8

with $\mathbf{x}_{t'}$9 implemented either as an EMA target or, in improved CT variants, as a stop-gradient copy of the current network (Liu et al., 2024, Song et al., 2023).

Later work expands this basic picture. Independent consistency training (ICT), used in plug-and-play MIMO channel estimation, applies a Pseudo-Huber loss across pairs of noise levels on a grid $$f_\theta(\mathbf{x}_t,t)=f_\theta(\mathbf{x}_{t'},t'),$$0, with boundary-condition preconditioning $$f_\theta(\mathbf{x}_t,t)=f_\theta(\mathbf{x}_{t'},t'),$$1 and a discretization size increasing from $$f_\theta(\mathbf{x}_t,t)=f_\theta(\mathbf{x}_{t'},t'),$$2 to $$f_\theta(\mathbf{x}_t,t)=f_\theta(\mathbf{x}_{t'},t'),$$3 (Li et al., 26 Apr 2026). In text-to-speech, CM-TTS combines a CT loss with reconstruction losses on mel, duration, pitch, and energy, and further modulates the sampling of time indices through uniform, linear, or importance-weighted samplers (Li et al., 2024).

Sampling is correspondingly simple. In the canonical 1-NFE regime, one draws $$f_\theta(\mathbf{x}_t,t)=f_\theta(\mathbf{x}_{t'},t'),$$4 and outputs $$f_\theta(\mathbf{x}_t,t)=f_\theta(\mathbf{x}_{t'},t'),$$5. Few-NFE sampling applies $$f_\theta(\mathbf{x}_t,t)=f_\theta(\mathbf{x}_{t'},t'),$$6 sequentially along a schedule of times, often with explicit re-noising between denoising steps. The original CM paper also emphasizes that the same multistep mechanism supports zero-shot editing, including inpainting, colorization, super-resolution, denoising, and interpolation, by inserting task-specific data-consistency operations between consistency evaluations (Song et al., 2023, Liu et al., 2024).

An important reinterpretation appears in Easy Consistency Tuning (ECT). ECT argues that diffusion training is a special case of consistency training obtained by choosing the largest possible interval $$f_\theta(\mathbf{x}_t,t)=f_\theta(\mathbf{x}_{t'},t'),$$7 in a finite-difference form of the consistency condition. On that reading, diffusion models correspond to a coarse enforcement of consistency, while CM fine-tuning progressively shrinks the interval so that $$f_\theta(\mathbf{x}_t,t)=f_\theta(\mathbf{x}_{t'},t'),$$8. This permits initializing from a pretrained diffusion model and tightening the consistency condition over training, rather than training a CM from scratch or distilling from a separate teacher at every step (Geng et al., 2024).

3. Architectural and algorithmic evolution

Subsequent work broadens the CM design space beyond endpoint projection. Consistency Trajectory Models (CTMs) replace the fixed terminal map by a function $$f_\theta(\mathbf{x}_t,t)=f_\theta(\mathbf{x}_{t'},t'),$$9 that can jump between arbitrary initial and final times along the PF ODE, while also exposing a score estimate at the $$f_\theta(\mathbf{x}_\epsilon,\epsilon)=\mathbf{x}_\epsilon$$0 limit. This restores access to standard diffusion guidance and likelihood computation, and empirically avoids the quality degradation that ordinary CM multistep samplers can exhibit as NFE increases. CTM reports single-step FID 1.73 on CIFAR-10 and 1.92 on ImageNet 64×64 (Kim et al., 2023).

Phased Consistency Models (PCMs) instead partition the PF ODE into sub-trajectories with edges $$f_\theta(\mathbf{x}_\epsilon,\epsilon)=\mathbf{x}_\epsilon$$1, and learn a separate phase function $$f_\theta(\mathbf{x}_\epsilon,\epsilon)=\mathbf{x}_\epsilon$$2 that maps any $$f_\theta(\mathbf{x}_\epsilon,\epsilon)=\mathbf{x}_\epsilon$$3 in $$f_\theta(\mathbf{x}_\epsilon,\epsilon)=\mathbf{x}_\epsilon$$4 to the boundary point $$f_\theta(\mathbf{x}_\epsilon,\epsilon)=\mathbf{x}_\epsilon$$5. This phase-wise decomposition eliminates the need to repeatedly jump to one global origin and thereby avoids the stochastic inconsistency that latent CMs exhibit across step counts. The same idea reappears in MotionPCM for text-conditioned human motion, where phased consistency training in latent space yields real-time inference at over 30 frames per second in a single sampling step, with 1-step/2-step/4-step FID values of 0.054, 0.040, and 0.036 on HumanML3D (Wang et al., 2024, Jiang et al., 31 Jan 2025).

A second line of development targets instability in continuous-time training. Flow-Anchored Consistency Models (FACM) argue that shortcut-only objectives lose track of the instantaneous velocity field and therefore mix a Flow Matching anchor with the CM shortcut objective without architectural changes. FACM reports ImageNet 256×256 FID 1.76 at NFE=1 and 1.32 at NFE=2 (Peng et al., 4 Jul 2025). Dual-End Consistency Model (DE-CM) makes a related diagnosis, attributing instability to an unstable self-supervised term and inflexible sampling to error accumulation. It decomposes the PF ODE into three sub-trajectory clusters—consistency, instantaneous, and noise-to-noisy—and reports one-step FID 1.70 and two-step FID 1.33 on ImageNet 256×256 (Dong et al., 11 Feb 2026).

Discretization and optimization have also become explicit design variables. Adaptive Discretization for Consistency Models (ADCM) formulates time-step selection as an optimization balancing local consistency for trainability and global consistency for stability, solving for $$f_\theta(\mathbf{x}_\epsilon,\epsilon)=\mathbf{x}_\epsilon$$6 by a Gauss-Newton update. On ImageNet 64×64, ADCM improves training efficiency with about 4% per-epoch overhead and strong 1-step FID gains (Bai et al., 20 Oct 2025). Orthogonally, LCSC shows that non-convex linear combinations of saved CM checkpoints can reach high-quality basins not accessed by SGD or EMA; for CM on CIFAR-10 CD it reports about 23× training speedup, while also improving 1-NFE quality (Liu et al., 2024).

4. Theory, robustness, and recurring failure modes

The first non-asymptotic convergence guarantees for CMs establish that, under $$f_\theta(\mathbf{x}_\epsilon,\epsilon)=\mathbf{x}_\epsilon$$7-accurate score matching, $$f_\theta(\mathbf{x}_\epsilon,\epsilon)=\mathbf{x}_\epsilon$$8-accurate consistency, Lipschitz score fields, and bounded second moment, one-step CM sampling can approximate the data distribution with small Wasserstein-2 error. The same analysis shows that Multistep Consistency Sampling can further reduce the error, and that simple OU smoothing or underdamped Langevin refinements yield total-variation guarantees. These results are notable because they do not require log-Sobolev inequalities or log-concavity assumptions (Lyu et al., 2023).

A separate theoretical issue is train–test mismatch. Diffusion models are trained on noisy states drawn from the forward process but sampled on states produced by the model itself, and the same mismatch propagates to CM distillation because the practical transport operator $$f_\theta(\mathbf{x}_\epsilon,\epsilon)=\mathbf{x}_\epsilon$$9 is only an approximation to the target PF ODE step. A distributionally robust optimization analysis shows that this mismatch can be mitigated by adversarial training. For Latent Consistency Models on MS-COCO 2014, adversarial training improves FID from 25.43 to 23.34 at 1 step, from 12.61 to 11.28 at 2 steps, and from 11.61 to 10.31 at 4 steps, while slightly improving CLIP score (Wang et al., 24 Feb 2025).

The empirical failure modes isolated in later work are relatively consistent. PCM identifies three flaws in latent CMs: stochastic multi-step inconsistency, CFG sensitivity combined with weak negative-prompt response, and poor very-low-step quality under pure $\mathbf{x}_\epsilon$0/Huber supervision (Wang et al., 2024). CCM finds that a ControlNet trained for diffusion can transfer high-level semantic control to CMs but struggles with low-level detail and realism, especially for sketch guidance and masked inpainting (Xiao et al., 2023). DE-CM attributes continuous-time divergence to an unstable self-supervised tangent term and multi-step inflexibility to accumulated sampling error, while ADCM formalizes the trainability–stability trade-off as a discretization problem rather than a purely heuristic scheduling issue (Dong et al., 11 Feb 2026, Bai et al., 20 Oct 2025).

These analyses collectively shift CM research away from the early view that few-step generation is mainly a distillation problem. They suggest instead that endpoint consistency, transport geometry, solver error, conditioning mechanisms, and schedule design are tightly coupled. A plausible implication is that later CM variants succeed not by abandoning the original self-consistency idea, but by embedding it in stronger trajectory structure.

5. Conditional, geometric, and constrained extensions

Conditional control in CMs has developed along three routes. CCM studies direct reuse of diffusion-trained ControlNet modules, CM-native ControlNets trained under consistency training, and a lightweight adapter that corrects diffusion ControlNet features for CM inference. The qualitative result is sharply asymmetric: direct transfer preserves high-level semantic structure, but CM-native control or adapter-based transfer is needed for realistic low-level detail, better mask fidelity, and stronger condition adherence (Xiao et al., 2023).

On non-Euclidean domains, the Riemannian Consistency Model (RCM) replaces Euclidean linearization by an exponential-map parameterization,

$\mathbf{x}_\epsilon$1

and replaces Euclidean distances by geodesic distances and ordinary derivatives by covariant derivatives. RCM derives both discrete- and continuous-time training objectives, proves equivalence between teacher-driven Riemannian consistency distillation and training-from-scratch Riemannian consistency training, and reports about $\mathbf{x}_\epsilon$2–$\mathbf{x}_\epsilon$3 speedup versus 100-step Riemannian Flow Matching in two-step sampling on flat tori, spheres, and $\mathbf{x}_\epsilon$4 (Cheng et al., 1 Oct 2025).

A more general abstraction appears in generalized CMs for constrained distribution matching. There, the diffusion-specific PF ODE is replaced by an arbitrary continuous normalizing flow path $\mathbf{x}_\epsilon$5, and the CM objective becomes a quadratic consistency loss along that path:

$\mathbf{x}_\epsilon$6

This formulation is used to treat latent-variable modeling, domain translation, and domain adaptation under explicit constraints on the mapping class, replacing GAN-style min–max objectives by single-level norm minimization (Shrestha et al., 17 Aug 2025).

Taken together, these extensions indicate that “consistency” is not restricted to Euclidean image denoising. It can be imposed on manifold-valued states, constrained pushforwards, phase-local latent trajectories, or explicitly conditioned inverse-problem solvers, provided the target transport admits a coherent time-indexed path.

6. Application domains and empirical profile

The reported application record of CMs spans communications, medical imaging, speech, motion, reinforcement learning, and inverse problems.

Domain	Reported result	Paper
MIMO channel estimation	4 PnP-CM iterations; NMSE reaches $\mathbf{x}_\epsilon$7 dB at 20 dB SNR with $\mathbf{x}_\epsilon$8	(Li et al., 26 Apr 2026)
Total-body PET posterior sampling	3 denoising steps; median absolute percent error $\mathbf{x}_\epsilon$9; median K-L divergence $\mathbf{x}_0$0; more than five orders of magnitude faster than MCMC	(Zhao et al., 17 Sep 2025)
Text-to-speech	On VCTK at $\mathbf{x}_0$1: MOS $\mathbf{x}_0$2, melFID $\mathbf{x}_0$3, RTF $\mathbf{x}_0$4	(Li et al., 2024)
Reinforcement learning policies	On hopper-medium-expert-v2: training time per epoch 57.06 s vs 31.94 s, inference 3.76 ms vs 1.84 ms for diffusion vs consistency policy	(Ding et al., 2023)
Inverse imaging	2-step measurement-guided CM sampling improves KID/FID across Fashion-MNIST and LSUN Bedroom tasks	(Tanevardi et al., 2 Oct 2025)
Motion synthesis	Over 30 FPS in one sampling step; 1-step FID 0.054 on HumanML3D	(Jiang et al., 31 Jan 2025)

The communications example is structurally notable because it embeds a pretrained CM prior in an explicit inverse-problem solver. In plug-and-play MIMO channel estimation, ADMM alternates a linear data-consistency solve with a CM denoiser acting on the angular-domain channel, and most of the NMSE improvement occurs within the first two to three outer iterations. The same study also shows a substantial out-of-distribution gap between the training scenario s002 and the test scenario s007, increasing from about 5.42 dB at $\mathbf{x}_0$5 dB SNR to about 10.60 dB at 20 dB SNR. This makes scenario dependence and adaptive scheduling central open issues rather than secondary engineering details.

The PET application illustrates a different use of CMs: posterior rather than unconditional generation. A conditional CM takes a voxel’s time-activity curve and arterial input function and generates posterior samples of kinetic parameters in only three denoising evaluations. On a full 4.4 million-voxel total-body PET volume, the reported runtime is 1.25 hours, whereas the MCMC reference would exceed one month of wall-clock time on a 128-core CPU cluster. The same framework also yields voxel-wise credible intervals and reversibility maps, not only point estimates.

Speech and reinforcement learning show that the few-step property is not tied to visual sampling. CM-TTS uses a consistency decoder for mel-spectrogram synthesis and combines it with weighted time samplers to improve speaker similarity, MOS, and WER at real-time latency. In reinforcement learning, consistency policies model multimodal action distributions while shortening both forward sampling chains and actor-gradient backpropagation paths, which is why the wall-clock gains persist in offline, offline-to-online, and fully online settings.

This application profile suggests that the main empirical attraction of CMs is neither one-step image generation alone nor diffusion distillation in isolation. It is the combination of an explicit trajectory-level prior, a short differentiable sampling path, and compatibility with domain structure such as linear measurement models, text conditions, speaker embeddings, kinetic simulators, or control states.