Phased Consistency Models: Deterministic Refinement

• Phased Consistency Models (PCM) are consistency-based generative frameworks that split the PF-ODE trajectory into separate phases, assigning unique target mappings within each phase.
• PCM improves low-step quality and controllability by mitigating noise accumulation and exposure artifacts through phase-local consistency and adversarial supervision.
• PCM demonstrates robust multi-step refinement in latent image and motion synthesis, achieving lower FID scores and real-time deterministic generation.

Phased Consistency Models (PCMs) are consistency-based generative models that partition the probability flow ODE (PF-ODE) trajectory of a diffusion process into multiple sub-trajectories, or phases, and enforce self-consistency only within each phase rather than across the entire path. In the original formulation, PCM was introduced to address limitations of Latent Consistency Models (LCMs) for high-resolution, text-conditioned latent image generation, especially under multi-step refinement and strong classifier-free guidance (CFG) (Wang et al., 2024). A subsequent motion-domain instantiation, MotionPCM, applied the same phase-wise principle to text-conditioned human motion synthesis in VAE latent space, combining phased consistency distillation with an EMA target student and an adversarial discriminator to obtain deterministic single- or few-step generation with real-time throughput (Jiang et al., 31 Jan 2025).

1. Origins and problem formulation

PCM is situated at the intersection of diffusion sampling, consistency distillation, and latent generative modeling. Diffusion models define a forward conditional path of the form $\alpha_t x_0 + \sigma_t \epsilon$, equivalently represented by a stochastic differential equation

$d x_t = f_t x_t\, d t + g_t\, d w_t,$

with a corresponding reverse-time PF-ODE

$$d x = \left[f_t - \frac{g_t^2}{2}\nabla_x \log P_t(x)\right] d t.$$

Standard sampling numerically solves this ODE using solvers such as DDIM, DPM-Solver, Euler, or Heun over $K$ steps; in few-step regimes, discretization error becomes pronounced (Wang et al., 2024).

Consistency Models (CMs) replace score estimation with a direct mapping $f_\theta(x,t)$ that sends any point on a trajectory to a common solution point $x_\epsilon$, enforcing

$$f_\theta(x_t,t)=x_\epsilon,\qquad f_\theta(x_t,t)=f_\theta(x_{t'},t'),\ \forall t,t'\in[\epsilon,T].$$

This eliminates long iterative denoising chains, but standard CM and LCM formulations retain important limitations. The image-domain PCM paper identifies three flaws in LCMs: inconsistency across inference step counts due to purely stochastic multi-step sampling, weak controllability under larger CFG and negative prompting, and poor few-step quality because traditional $L_2$ or Huber losses provide insufficient fine-grain supervision (Wang et al., 2024). The motion-domain paper reports closely related failures when CM or latent CM is applied to text-conditioned motion in latent space: stochastic noise accumulation across multiple steps and degraded quality in low-step settings (Jiang et al., 31 Jan 2025).

The central PCM intervention is therefore not merely to reduce step count, but to change the consistency target itself. Rather than forcing the entire PF-ODE trajectory to collapse to a single early-time solution, PCM assigns a separate consistency target to each sub-trajectory. This suggests that PCM should be understood as a redefinition of the consistency manifold for latent diffusion, not only as a faster sampler.

2. Phase-wise consistency formulation

PCM splits the PF-ODE trajectory into $M$ phases with edge timesteps

$s_0=\epsilon < s_1 < \cdots < s_M=T.$

For each sub-trajectory $d x_t = f_t x_t\, d t + g_t\, d w_t,$0, PCM defines an independent consistency function

$d x_t = f_t x_t\, d t + g_t\, d w_t,$1

and enforces phase-local self-consistency: $d x_t = f_t x_t\, d t + g_t\, d w_t,$2 In MotionPCM the same idea is written as

$d x_t = f_t x_t\, d t + g_t\, d w_t,$3

with the explicit interpretation that any noisy point within a phase maps to the start of that phase rather than to $d x_t = f_t x_t\, d t + g_t\, d w_t,$4 (Jiang et al., 31 Jan 2025).

This phase decomposition enables deterministic multi-step refinement across phase boundaries. The composition is defined as

$d x_t = f_t x_t\, d t + g_t\, d w_t,$5

which maps a point from phase $d x_t = f_t x_t\, d t + g_t\, d w_t,$6 to the solution point of phase $d x_t = f_t x_t\, d t + g_t\, d w_t,$7 (Wang et al., 2024). MotionPCM uses the same transition-map principle during inference, sampling $d x_t = f_t x_t\, d t + g_t\, d w_t,$8, composing phase transitions to obtain $d x_t = f_t x_t\, d t + g_t\, d w_t,$9, and decoding with the VAE decoder (Jiang et al., 31 Jan 2025).

The phase-local mapping is parameterized by a teacher-informed exact-solution form. Using the PF-ODE solution from $$d x = \left[f_t - \frac{g_t^2}{2}\nabla_x \log P_t(x)\right] d t.$$0 to $$d x = \left[f_t - \frac{g_t^2}{2}\nabla_x \log P_t(x)\right] d t.$$1,

$$d x = \left[f_t - \frac{g_t^2}{2}\nabla_x \log P_t(x)\right] d t.$$2

PCM introduces

$$d x = \left[f_t - \frac{g_t^2}{2}\nabla_x \log P_t(x)\right] d t.$$3

and sets

$$d x = \left[f_t - \frac{g_t^2}{2}\nabla_x \log P_t(x)\right] d t.$$4

The paper proves that this parameterization is equivalent to a DDIM-form update,

$$d x = \left[f_t - \frac{g_t^2}{2}\nabla_x \log P_t(x)\right] d t.$$5

establishing a direct connection between phased consistency prediction and deterministic ODE sampling (Wang et al., 2024).

The original PCM work further proves phase-wise and cross-phase approximation bounds of order $$d x = \left[f_t - \frac{g_t^2}{2}\nabla_x \log P_t(x)\right] d t.$$6 under Lipschitz and solver-error assumptions. These results formalize the intuition that phase partitioning reduces accumulation error by shortening the consistency horizon (Wang et al., 2024).

3. Distillation objective, adversarial supervision, and computational structure

PCM is trained by teacher–student distillation. A teacher diffusion model $$d x = \left[f_t - \frac{g_t^2}{2}\nabla_x \log P_t(x)\right] d t.$$7 provides ODE targets through a $$d x = \left[f_t - \frac{g_t^2}{2}\nabla_x \log P_t(x)\right] d t.$$8-step solver $$d x = \left[f_t - \frac{g_t^2}{2}\nabla_x \log P_t(x)\right] d t.$$9, typically DDIM. The student $K$0 predicts a phase-start solution directly from a noisier point in the same phase, while an EMA target network $K$1 stabilizes training. In the image formulation, the phased consistency loss is defined over uniformly sampled phases and intervals, with

$K$2

followed by a distance term between the online and EMA phase mappings (Wang et al., 2024).

MotionPCM adopts the same structure in VAE latent space. Its phased consistency distillation loss is written as

$K$3

where $K$4 is the Huber distance, the teacher is a pre-trained latent diffusion model from MLD, the online student is initialized from the teacher, and the target student is updated by

$K$5

with $K$6 in the reported motion experiments (Jiang et al., 31 Jan 2025).

A distinctive PCM component is adversarial latent supervision for low-step regimes. In MotionPCM, the discriminator $K$7 is conditioned on timestep $K$8 and text $K$9, and the hinge-style adversarial loss is

$f_\theta(x,t)$0

with total objective

$f_\theta(x,t)$1

The general PCM paper likewise reports that adversarial consistency loss improves low-step FID and stabilizes visuals, that a latent-space discriminator is superior to a DINO-based pixel discriminator, and that replacing this with a normal GAN loss causes unstable training and conflicts with consistency distillation (Wang et al., 2024, Jiang et al., 31 Jan 2025).

From a complexity perspective, the PCM paper contrasts diffusion and CM objectives, which scale as $f_\theta(x,t)$2, with CTM objectives at $f_\theta(x,t)$3. PCM preserves $f_\theta(x,t)$4 training while introducing multiple phase-specific functions and an EMA target, and its inference becomes deterministic $f_\theta(x,t)$5-step refinement across phase edges, with teacher-solver cost incurred only during distillation (Wang et al., 2024). MotionPCM expresses the same efficiency gain operationally: diffusion sampling is $f_\theta(x,t)$6 forward passes for $f_\theta(x,t)$7 steps, whereas MotionPCM is $f_\theta(x,t)$8 for single-step or $f_\theta(x,t)$9 for a few steps with $x_\epsilon$0, plus one VAE decode (Jiang et al., 31 Jan 2025).

4. Conditioning, controllability, and inference dynamics

One of PCM’s main claims is that phase-wise consistency improves controllability in latent text-conditioned generation. In the image formulation, CFG is written as

$x_\epsilon$1

The paper argues that guided distillation in LCMs causes exposure artifacts and weak negative-prompt sensitivity; its theorem on CFG shows that the student under guided distillation inherits multiplicative scaling $x_\epsilon$2 and a merged negative/null prediction, which explains why large CFG values and negative prompts behave poorly in prior designs (Wang et al., 2024).

PCM therefore allows a variant in which the teacher solver omits CFG augmentation during distillation (“Ours*”). The reported effect is improved prompt alignment and higher sensitivity to negative prompts. The paper also states that PCM weights support robust multi-step refinement under larger CFG ranges than LCMs, whereas LCMs can only accept CFG with 1–2 steps and become exposed at larger values (Wang et al., 2024).

Inference can be purely deterministic or partially stochastic. Deterministic refinement composes the phase mappings from $x_\epsilon$3 to $x_\epsilon$4, for example $x_\epsilon$5 in a two-phase setting. PCM also introduces an optional stochastic interpolation

$x_\epsilon$6

where $x_\epsilon$7, $x_\epsilon$8 is purely deterministic, and $x_\epsilon$9 recovers fully stochastic CM sampling (Wang et al., 2024).

The image paper further notes a numerical instability at 1-step inference caused by very small $$f_\theta(x_t,t)=x_\epsilon,\qquad f_\theta(x_t,t)=f_\theta(x_{t'},t'),\ \forall t,t'\in[\epsilon,T].$$0 in Stable Diffusion latents, with $$f_\theta(x_t,t)=x_\epsilon,\qquad f_\theta(x_t,t)=f_\theta(x_{t'},t'),\ \forall t,t'\in[\epsilon,T].$$1 for SD(1.5/XL). To mitigate large scaling when predicting $$f_\theta(x_t,t)=x_\epsilon,\qquad f_\theta(x_t,t)=f_\theta(x_{t'},t'),\ \forall t,t'\in[\epsilon,T].$$2, it clips the denominator: $$f_\theta(x_t,t)=x_\epsilon,\qquad f_\theta(x_t,t)=f_\theta(x_{t'},t'),\ \forall t,t'\in[\epsilon,T].$$3 This clarifies that PCM’s one-step setting is not treated as trivial; it requires explicit stabilization and, in the paper’s own wording, can still be unstable, producing structural errors or blurry images when the number of steps is very low (Wang et al., 2024).

5. Results in image and video generation

The original PCM paper evaluates image models on COCO-2014 Karpathy 30K and CC12M-30K, with training on CC3M, using SD v1.5 and SDXL backbones; it evaluates video generation on UCF-101 captions after training on WebVid-2M via AnimateLCM-style temporal inflation (Wang et al., 2024).

Domain	Setting	Reported PCM outcome
SD v1.5 image generation	COCO-30K, 1 step	FID 17.91; FID-CLIP 11.66; FID-SD 8.27; CLIP 29.26; CONSISTENCY 0.81
SDXL image generation	COCO-30K, 1 step	FID 21.23; FID-CLIP 13.66; FID-SD 9.32; CLIP 31.55; CONSISTENCY 0.83
Video generation	UCF-101 captions, 1/2/4 steps	CLIP Score 29.88 / 30.22 / 30.72
Motion generation	HumanML3D, 1 step	AITS 0.031 s; over 30 FPS; FID 0.054

For SD v1.5 multi-step image generation, the reported FID-SD on COCO-30K is $$f_\theta(x_t,t)=x_\epsilon,\qquad f_\theta(x_t,t)=f_\theta(x_{t'},t'),\ \forall t,t'\in[\epsilon,T].$$4 for PCM across 1, 2, 4, 8, and 16 steps, and on CC12M-30K it is $$f_\theta(x_t,t)=x_\epsilon,\qquad f_\theta(x_t,t)=f_\theta(x_{t'},t'),\ \forall t,t'\in[\epsilon,T].$$5 (Wang et al., 2024). The paper states that PCM outperforms LCM and CTM at most step counts, and that the performance gap grows with the number of refinement steps, which it interprets as evidence for better multi-step refinement. For SDXL on CC12M-30K, the reported FID-SD values are $$f_\theta(x_t,t)=x_\epsilon,\qquad f_\theta(x_t,t)=f_\theta(x_{t'},t'),\ \forall t,t'\in[\epsilon,T].$$6, while SDXL-LCM and SDXL-CTM degrade substantially at some step counts (Wang et al., 2024).

In video generation, PCM extends the same phased consistency objective to temporally inflated SD v1.5 latents. On UCF-101 captions, PCM reports CLIP Score $$f_\theta(x_t,t)=x_\epsilon,\qquad f_\theta(x_t,t)=f_\theta(x_{t'},t'),\ \forall t,t'\in[\epsilon,T].$$7, Flow Magnitude $$f_\theta(x_t,t)=x_\epsilon,\qquad f_\theta(x_t,t)=f_\theta(x_{t'},t'),\ \forall t,t'\in[\epsilon,T].$$8, and CLIP Consistency $$f_\theta(x_t,t)=x_\epsilon,\qquad f_\theta(x_t,t)=f_\theta(x_{t'},t'),\ \forall t,t'\in[\epsilon,T].$$9 for 1-, 2-, and 4-step sampling, compared with AnimateLCM’s lower CLIP Score and substantially lower Flow Magnitude (Wang et al., 2024). The paper characterizes this as state-of-the-art few-step text-to-video generation.

These results support two empirical interpretations. First, PCM’s primary strength is not limited to one-step sampling; the design is explicitly optimized for multi-step deterministic refinement. Second, the latent adversarial signal appears especially important when VAE decoding magnifies low-step prediction errors.

6. MotionPCM and real-time human motion synthesis

MotionPCM transfers PCM from image/video latents to text-conditioned human motion synthesis. It adopts a two-stage latent pipeline consisting of a pre-trained VAE backbone, a teacher latent diffusion model, phased student networks, and a discriminator (Jiang et al., 31 Jan 2025). The encoder $L_2$0 maps motion sequences $L_2$1 to latents $L_2$2, and the decoder $L_2$3 reconstructs motion. The VAE incorporates MotionLCM-V2 improvements, specifically a trainable linear layer after latent tokens to strengthen multimodal modulation and the removal of unnecessary ReLUs to preserve negative components in text features.

The teacher is an epsilon-prediction latent diffusion model trained in VAE latent space following MLD. The online PCM student $L_2$4 is initialized from this teacher and trained to map $L_2$5 in phase $L_2$6 directly to $L_2$7; the target student $L_2$8 is updated by EMA; the discriminator $L_2$9 distinguishes phase-noised student latents from data-driven latents while conditioning on timestep and text. DDIM serves as the deterministic ODE solver for skip-solved teacher targets. The end-to-end training pipeline freezes the VAE encoder and teacher, trains the online PCM and discriminator, and updates the target PCM by EMA (Jiang et al., 31 Jan 2025).

MotionPCM sets the number of phases $M$0 equal to the desired number of inference steps $M$1, with edge times $M$2 and $M$3. For each phase, the teacher skip-solves from a noisier latent $M$4 to $M$5, the target network maps this to $M$6, and the online network predicts $M$7 directly from the noisier point. The paper emphasizes that this avoids reintroducing stochastic noise at intermediate steps and yields deterministic few-step sampling (Jiang et al., 31 Jan 2025).

The reported experiments use HumanML3D, with 14,616 motions and 44,970 text annotations, and motion features including root velocity, root height, local joint positions and velocities, root-space rotations, and foot-contact indicators (Jiang et al., 31 Jan 2025). Evaluation uses AITS, R-Precision (Top-1/2/3), FID, Multimodal Distance, Diversity, and MultiModality against a broad baseline set including MLD, MotionLCM-V1/V2, MDM, MotionDiffuse, ReMoDiffuse, MoMask, StableMoFusion, and others.

The principal quantitative results are strongly few-step oriented. MotionPCM reports AITS of $M$8 s, $M$9 s, and $s_0=\epsilon < s_1 < \cdots < s_M=T.$0 s for 1-, 2-, and 4-step inference, respectively; the paper notes that the 1-step setting supports motion synthesis over 30 frames per second (Jiang et al., 31 Jan 2025). FID is reported as $s_0=\epsilon < s_1 < \cdots < s_M=T.$1 at 1 step, $s_0=\epsilon < s_1 < \cdots < s_M=T.$2 at 2 steps, and $s_0=\epsilon < s_1 < \cdots < s_M=T.$3 at 4 steps, with the best 2-step R-Precision values $s_0=\epsilon < s_1 < \cdots < s_M=T.$4 for Top-1/2/3 and best 2-step MM Dist $s_0=\epsilon < s_1 < \cdots < s_M=T.$5. Relative to MotionLCM-V2 at the same step count, the paper reports FID reductions of $s_0=\epsilon < s_1 < \cdots < s_M=T.$6 at 1 step ($s_0=\epsilon < s_1 < \cdots < s_M=T.$7), $s_0=\epsilon < s_1 < \cdots < s_M=T.$8 at 2 steps ($s_0=\epsilon < s_1 < \cdots < s_M=T.$9), and $d x_t = f_t x_t\, d t + g_t\, d w_t,$00 at 4 steps ($d x_t = f_t x_t\, d t + g_t\, d w_t,$01), along with consistent R-Precision gains and a best 2-step MM Dist improvement from $d x_t = f_t x_t\, d t + g_t\, d w_t,$02 to $d x_t = f_t x_t\, d t + g_t\, d w_t,$03 (Jiang et al., 31 Jan 2025). The abstract additionally states a $d x_t = f_t x_t\, d t + g_t\, d w_t,$04 FID improvement over the previous state of the art.

Ablation results indicate that teacher skip length, EMA, and adversarial supervision all matter. Reducing DDIM skip from $d x_t = f_t x_t\, d t + g_t\, d w_t,$05 to $d x_t = f_t x_t\, d t + g_t\, d w_t,$06 worsens most metrics, with FID degrading from $d x_t = f_t x_t\, d t + g_t\, d w_t,$07 to $d x_t = f_t x_t\, d t + g_t\, d w_t,$08; removing EMA slightly worsens performance; removing the discriminator dramatically harms FID and MM Dist (Jiang et al., 31 Jan 2025). Qualitatively, MotionPCM is reported to better follow complex prompts involving turning around and returning to the start, limp walking, or walking while raising a hand and circling, relative to MotionLCM-V2 and MLD.

7. Limitations, positioning, and open directions

PCM is positioned in the literature as a generalization of CM and LCM rather than a direct replacement for diffusion modeling. Standard CM learns a global consistency target $d x_t = f_t x_t\, d t + g_t\, d w_t,$09; LCM transfers this idea to latent space; PCM redefines the target phase-wise, enabling deterministic multi-step refinement and stronger low-step behavior (Wang et al., 2024). MotionPCM further shows that the same principle can be integrated into a motion-specific latent pipeline with text conditioning and adversarial guidance (Jiang et al., 31 Jan 2025).

A common misconception is that PCM is simply “LCM with more steps.” The published formulations indicate otherwise. PCM changes the structure of the learning objective, the inference path, and the teacher–student relationship by introducing phase-local targets, compositional phase transitions, and exact-solution parameterization through $d x_t = f_t x_t\, d t + g_t\, d w_t,$10 (Wang et al., 2024). Another misconception is that PCM is image-specific; the available results cover images, video, and text-conditioned human motion (Wang et al., 2024, Jiang et al., 31 Jan 2025).

The limitations are explicit. In the original PCM paper, very low-step generation, especially one-step generation, can still be unstable and may produce structural errors or blurry images, although adversarial latent loss and multi-step refinement mitigate this (Wang et al., 2024). In MotionPCM, MultiModality is described as competitive but not state of the art, Diversity is second-best at 4 steps rather than best, and evaluation is limited to HumanML3D, with future work proposed on additional datasets to assess generalizability (Jiang et al., 31 Jan 2025). The motion paper also notes that broader stress tests for rare actions and very long sequences are not exhaustively reported.

The current evidence therefore supports a narrower but technically clear conclusion: PCM is a phase-wise reformulation of consistency distillation at the PF-ODE level that improves deterministic few-step latent generation, especially when conventional LCM training suffers from stochastic step inconsistency, CFG fragility, or low-step quality loss. MotionPCM demonstrates that the same reformulation remains effective when the latent variable is not an image code but a structured motion representation, yielding real-time text-conditioned synthesis with improved alignment and plausibility (Wang et al., 2024, Jiang et al., 31 Jan 2025).