Skip to content

Dual-End Consistency Model

Conference: ECCV 2026
arXiv: 2602.10764
Code: None
Area: Diffusion Models
Keywords: Consistency Models, Diffusion Distillation, Few-step Generation, Flow Matching, PF-ODE

TL;DR

DE-CM achieves SOTA performance on ImageNet 256 with a 1-step generation FID of 1.70. It optimizes three key sub-trajectories (consistency trajectory, instantaneous velocity trajectory, and noise-to-noise trajectory) selected from the entire PF-ODE path, stabilizes training with flow matching as boundary regularization, and mitigates error accumulation via an N2N mapping.

Background & Motivation

Diffusion models and flow matching models have made significant progress in tasks such as image, video, and 3D generation. However, their slow iterative sampling remains a core bottleneck for deployment. Consistency Models (CMs) have emerged as one of the most effective distillation pathways for few-step generation by learning self-consistent functions that map noise directly to data along the PF-ODE trajectory. While the transition from discrete-time to continuous-time CMs eliminated discretization errors, it failed to address two fundamental issues: training instability and inflexible sampling.

Prior works have attempted to alleviate these issues through three directions: diffusion modeling (TrigFlow in sCMs, flow mapping in MeanFlow/AYF), neural network architecture adjustments (tailored designs), and training objective regularization (IMM, MeanFlow, SplitMeanFlow, sCoT, FACM). However, these methods overlook a critical insight: the core problem lies not in how to optimize, but in which trajectories to optimize. MeanFlow and AYF perform full enumeration optimization over the entire trajectory space (\(O(n^2)\) complexity), where the entangled learning objectives lead to slow convergence or even training collapse. Conversely, sCMs optimize only a single CM trajectory, where the lack of critical path constraints increases instability. FACM achieves better training stability by explicitly decoupling the instantaneous velocity target on the PF-ODE trajectory, indicating that selecting an appropriate subset of candidate trajectories is the key to simultaneously achieving training stability and high performance.

Key Insight: Analyze the root causes of instability and inflexibility in CMs, and selectively pick the most critical trajectory segments for optimization based on these findings. Core Idea: From the entire space of the PF-ODE trajectory \(\{(t, s) \mid t < s\}\), select only three key sub-trajectory clustersโ€”the consistency trajectory (\(s=1\)), the instantaneous velocity trajectory (\(s=t\)), and the noise-to-noise trajectory (\(r \to 0\))โ€”as optimization targets. Flow matching is employed as boundary regularization, and N2N mapping is used to eliminate error accumulation from long jumps.

Method

Overall Architecture

The core problem DE-CM aims to solve is the training instability and inflexible sampling of continuous-time CMs. The overall mechanism is to decompose the entire PF-ODE trajectory using an intermediate time \(t\) as a boundary, select only three sub-trajectory clusters critical for distillation as optimization targets, and handle them with three matching loss functions. The inputs are noise \(z\) and reference image/text conditions, and the output is the distilled model \(F_\theta\), supporting flexible sampling from 1 to 50 NFEs.

The framework utilizes a flow mapping parameterization \(f_\theta(x_t, t, s) = x_t + (s-t)F_\theta(x_t, t, s)\), where \(F_\theta\) is a neural network (built by adding the right-endpoint time condition \(s\) as input to a pre-trained LightningDiT), and \(t\) and \(s\) represent the left and right time endpoints of the trajectory segment, respectively. During training, three losses are computed simultaneously for each batch: the consistency distillation loss \(L_\text{cm}\) (\(s=1\), learning the mapping from noise to data), the flow matching loss \(L_\text{fm}\) (\(s=t\), acting as left boundary regularization to stabilize training), and the noise-to-noise mapping loss \(L_\text{n2n}\) (left endpoint \(r \to 0\), learning the mapping from pure noise to any intermediate state). Their sum is used to update the parameters.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Input: Noise z + Reference image z_ref"] --> B["Teacher Model F_tea<br/>Compute CFG velocity v_t"]
    B --> C["Consistency Trajectory (s=1)<br/>Continuous-Time CD Loss"]
    B --> D["Instantaneous Velocity Trajectory (s=t)<br/>Flow Matching Boundary Regularization"]
    B --> E["Noise-to-Noise Trajectory (rโ†’0)<br/>JVP Calculates N2N Target"]
    C --> F["L_cm + L_fm + L_n2n"]
    D --> F
    E --> F
    F --> G["Update Online Model F_ฮธ<br/>+ EMA Update F_ema"]

Key Designs

1. Trajectory Cluster Selection: From Full Enumeration to Three Critical Sub-trajectories

The root cause of training instability in CMs can be understood by decomposing their continuous-time objectives. Expanding the gradient objective of Eq. (6), the training objective of CMs is equivalent to the sum of two terms (Eq. 8): a fully supervised term \(\|F_\theta(x_t, t) - v_\phi\|^2\) (aligning with the instantaneous velocity of the teacher model) and a self-supervised term \(\|F_\theta(x_t, t) - \tilde{v}\|^2\) (aligning with the time-varying derivative correction of the current network state, where \(\tilde{v} = F_\theta^-(x_t, t) + (1-t) \cdot dF_\theta^-/dt\)). Experiments reveal that the gradient of the self-supervised term is much more unstable and fluctuates heavier than that of the fully supervised term. This is because before network convergence, the Jacobian-Vector Product computation of \(dF_\theta^-/dt\) is highly dependent on the current parameter state; a single erroneous parameter update can cause the online model to diverge. At the same time, overemphasizing consistency causes the model to "forget" the concept of instantaneous velocity learned during the pre-training stage (catastrophic forgetting).

Based on this analysis, instead of performing a full \(\{(t, s) \mid t < s\}\) space enumeration (as in MeanFlow/AYF, which has \(O(n^2)\) complexity and slows down convergence with redundant trajectories), DE-CM selects only three trajectory clusters with clear functional boundaries: - Consistency Trajectory (\(t \in [0, 1], s=1\)): Provides self-consistency constraints to enable few-step inference capabilities. - Instantaneous Velocity Trajectory (\(s=t\)): At this point, the self-supervised term is zero, leaving only the fully supervised flow matching objective, which serves as left boundary regularization to eliminate training instability. - Noise-to-Noise Trajectory (\(r \to 0, t\)): Learns the mapping from pure noise to any intermediate point, providing noise-awareness for the first step of inference and mitigating error accumulation from long jumps.

Experiments prove that these three trajectories are sufficientโ€”the FM + CD + N2N combination outperforms any subset combination across 1, 2, 4, and 50 NFEs.

2. N2N Mapping: Breaking the Restriction of Only Predicting \(x_1\)

Standard CMs are restricted by their self-consistent functions and the boundary condition \(f_\theta(x_1, 1) = x_1\), meaning they can only learn "long-jump" mappings from any \(x_t\) to data \(x_1\). This introduces three issues: fixed inference steps (restricted to long-jump sampling with \(\gamma=1\)), error accumulation (multiple long jumps cause the total variation distance to aggregate to \(O(\sqrt{T+t_1+\dots+t_N})\)), and sensitivity to initial noise.

The core idea of N2N mapping is to replace the left endpoint with \(r \to 0\) (i.e., pure noise \(x_0\)) instead of \(t\), and optimize the interval \([r, t]\) (\(r \to 0\)), allowing the model to learn the mapping from pure noise \(x_0\) to any intermediate state \(x_t\) rather than just data \(x_1\). Regarding the training objective, the discrete form of N2N is given by Eq. (13): \(\|f_\theta(x_r, r, t) - S_\psi(f_\theta^-(x_r, r, t'))\|^2\), where \(S_\psi\) is a single-step Euler solver along the PF-ODE from \(x_{t'}\) to \(x_t\). Conducting continuous-time limit derivations on both the right endpoint \(t\) and the left endpoint \(r\) (taking \(\epsilon \to 0\)), we obtain the unified gradient objective (Eq. 15):

\[\nabla_\theta \mathbb{E}[ w'(r,t) \cdot f_\theta^\top(x_r, r, t) \cdot ( \lambda \cdot \frac{df_\theta^-}{dt} - \gamma \cdot \frac{df_\theta^-}{dr} - \lambda \cdot v_\text{tea}(f_\theta^-(x_r, r, t), t) ) ]\]

where \(\lambda\) and \(\gamma\) are weighting coefficients that control the gradient contributions of the right endpoint \(t\) and left endpoint \(r\), respectively. Eventually, \(L_\text{n2n}\) is formulated into a practical form (Eq. 16): \(\|F_\theta - F_\theta^- + w'(r,t) \cdot g\|^2\), where \(g = (\lambda+\gamma)F_\theta^- - (\gamma \cdot v_\phi + \lambda \cdot v_\psi) - (t-r) \cdot \dot{F}_\theta^-\), and \(\dot{F}_\theta^-\) is the time-varying derivative computed via JVP as \([\nabla_{x_r}F_\theta^-, \partial F_\theta^-/\partial r, \partial F_\theta^-/\partial t] \cdot [\lambda \cdot v_\phi, \lambda, -\gamma]^\top\). In ablation studies, \(\lambda=0.5, \gamma=1\) achieves optimal results; a too small \(\gamma\) leads to optimization difficulties, while a too large or zero \(\lambda\) degrades performance.

3. Flow Matching Boundary Regularization: Filling the Blank Left by Self-Supervised Instability

Eq. (8) demonstrates that the CM objective consists of two MSE terms: \(\|F_\theta - v_\phi\|^2 + \|F_\theta - \tilde{v}\|^2\). When \(s \to t\), the second term becomes zero, leaving only the fully supervised first term, which is the standard flow matching loss. This observation reveals the true role of FM: instead of being an independent "auxiliary training target," it serves as a left boundary condition that provides deterministic anchors for optimization when the self-supervised term is unstable.

The designed \(L_\text{fm}\) (Eq. 18) not only contains the \(L_2\) loss \(\|F_\theta(x_t, t, s) - v_\phi\|^2\) (\(s=t\)), but also introduces a cosine similarity loss \(L_\text{cos}(F_\theta, v_\phi)\) to preserve the directional features of the pre-trained model, supporting multi-step ODE sampling without performance drops. Experiments validate the boundary regularization effect of FM: adding FM smoothens and stabilizes the gradient norm curve (Fig. 5a), preventing training collapse. In the ablation study, removing FM and relying solely on CM distillation (CD) degrades the 1-NFE FID from 1.70 to 3.08, and causes the high-NFE performance to collapse from 1.26 to 6.52. This indicates that without FM acting as a boundary constraint, the consistency self-supervised term not only makes training unstable but also leads to multi-step performance collapse, i.e., "worse performance with more distillation steps."

4. \(r\)-value Decay and Velocity Normalization: Two Key Training Techniques

In N2N training, the \(r\)-value should theoretically approach 0, but experiments show that setting \(r=0\) directly at the start of training causes severe gradient oscillations. DE-CM introduces an \(r\)-value decay strategy: \(r = \max(0, (1 - i/i_\text{max}) \cdot \delta)\), where \(\delta = 0.1\) and \(i_\text{max} = 20000\). During the first 20k iterations, \(r\) gradually decays from 0.1 to 0, providing a smooth transition period for the model. Ablation studies show that setting \(\delta = 0\) (no decay) yields a 1-NFE FID that is 0.04 worse than \(\delta = 0.1\).

Velocity normalization targets the oversaturation issue caused by CFG guidance: \(v_\phi^\text{cfg} = v_\phi^\text{cond} + (w_\text{cfg} - 1) \cdot \min(1, \eta / \|\Delta v\|) \cdot \Delta v\), where \(\Delta v = v_\phi^\text{cond} - v_\phi^\text{uncond}\). By bounding the norm of the conditional-unconditional velocity difference \(\Delta v\) with an upper limit \(\eta\), it prevents the oversaturation of generated images caused by CFG amplification. This also implicitly regularizes the JVP vectors, indirectly enhancing training stability. Removing velocity normalization in the ablation study increases the 1-NFE FID from 1.70 to 1.75.

A Complete Example: A Single Training Iteration

Taking a batch from class-to-image (C2I) training as an example, with batch size = 1 and \(w_\text{cfg} = 1.75\):

  1. Sample a reference image \(z_\text{ref}\) and category label \(y_\text{ref}\) from the dataset; sample noise \(z \sim \mathcal{N}(0, I)\) and time \(t \sim U(0, 1)\). Compute \(r\) based on the decay strategy.
  2. Construct perturbed samples: \(z_t = (1-t) \cdot z + t \cdot z_\text{ref}\); \(z_r = (1-r) \cdot z + r \cdot z_\text{ref}\).
  3. The teacher model computes the CFG velocity: \(v_t^\text{cond} = F_\text{tea}(z_t, t; y_\text{ref})\), \(v_t^\text{uncond} = F_\text{tea}(z_t, t; \emptyset)\), \(v_t = v_t^\text{uncond} + 1.75 \cdot (v_t^\text{cond} - v_t^\text{uncond})\) (normalized); compute \(v_r\) analogously.
  4. Parallel computation of three losses:
    • CD Loss: \(L_\text{cm} = \|F_\theta(z_t, t, 1; y_\text{ref}) - v_\text{cm}^\text{tar}\|^2\), where \(v_\text{cm}^\text{tar}\) is a weighted combination of \(v_\phi\) and \(F_\theta^-\).
    • FM Loss: \(L_\text{fm} = \|F_\theta(z_t, t, t; y_\text{ref}) - v_t\|^2 + L_\text{cos}(F_\theta, v_t)\).
    • N2N Loss: Compute the time-varying derivative \(\dot{F}_\theta^-\) of \(F_\theta(x_r, r, t; y_\text{ref})\) via JVP, construct \(g_\text{n2n}\) and \(v_\text{n2n}^\text{tar}\), and compute \(L_\text{n2n} = \|F_\text{n2n} - v_\text{n2n}^\text{tar}\|^2\).
  5. Total loss \(L = L_\text{cm} + L_\text{fm} + L_\text{n2n}\). Update the online model \(F_\theta\) and the EMA model \(F_\text{ema}\).

Loss & Training

The total loss is the sum of three parts: \(L = L_\text{cm} + L_\text{fm} + L_\text{n2n}\). The C2I task uses LightningDiT-XL/1 as the teacher model (675M parameters) with a latent space of \(32 \times 16 \times 16\), optimized via AdamW with \(lr = 1\text{e-}4\) for 250 epochs. The T2I task uses the SD3.5-Medium teacher model with a LoRA rank of 64, \(lr = 5\text{e-}4\), trained on 100K samples from text-to-image-2M. Key hyperparameters: \(\lambda=0.5, \gamma=1, \delta=0.1\), N2N update frequency \(\text{freq}=3\) (updating the N2N loss every 3 steps), and timesteps are sampled using uniform piecewise sampling rather than a log-normal distribution. The CFG weight is determined via a grid search on \(\eta\) that optimizes the teacher model's FID.

Key Experimental Results

Main Results

C2I: ImageNet \(256 \times 256\) Class-Conditional Generation FID Comparison (Table 1 Summary):

Method NFE Parameters FIDโ†“
GigaGAN 1 569M 3.45
StyleGAN-XL 1 166M 2.30
VAR-d30 10 2B 1.92
LightningDiT-XL/1 (w=6.7) 250ร—2 675M 1.35
MeanFlow 1 / 2 675M 3.43 / 2.20
MeanFlow-โ€  (250 epoch repro) 1 / 2 675M 2.79 / 1.74
FACM-โ€  (250 epoch repro) 1 675M 1.76
sCMs-โ€  (250 epoch repro) 1 / 2 675M 3.34 / 1.94
DE-CM (Ours) 1 / 2 / 50 675M 1.70 / 1.33 / 1.26

DE-CM outperforms all other distillation methods of the same scale with a 1-NFE FID of 1.70. At 2 NFEs (1.33), it even surpasses the teacher model LightningDiT at 250 NFEs (1.35). At 50 NFEs, the FID drops to 1.26, indicating that the model is not only strong in few-step generation but also avoids performance degradation at higher step counts.

T2I: Text-to-Image Generation Feedback Metrics Comparison (Partial Table 2 Summary):

Method NFE CLIPโ†‘ BLIPโ†‘ ImageRewardโ†‘
SD3.5-Medium 50ร—2 0.2998 0.5451 0.8750
Flux.1-Dev 50ร—2 0.2941 0.5384 0.9634
LCM 1 / 4 / 8 0.2508 / 0.2948 / 0.2993 0.4560 / 0.5306 / 0.5318 โˆ’1.0622 / 0.4910 / 0.5871
Hyper-SD 1 / 4 / 8 0.2939 / 0.2949 / 0.3029 0.5161 / 0.5307 / 0.5293 0.5704 / 0.7931 / 0.8447
DE-CM (Ours) 1 / 4 / 8 / 50 0.2996 / 0.2999 / 0.3000 / 0.2993 0.5398 / 0.5474 / 0.5479 / 0.5490 0.5758 / 0.8117 / 0.8671 / 0.9712

For T2I, DE-CM achieves the best BLIP scores across all NFE settings. At 1 NFE, its ImageReward score of 0.5758 significantly outperforms other 1-NFE methods like LCM (-1.0622) and CTM (-0.7639). At a high NFE of 50, both its BLIP score (0.5490) and ImageReward score (0.9712) reach levels comparable to or slightly better than multi-step pre-trained large models (Flux, SD3.5).

Ablation Study

Configuration NFE=1 NFE=2 NFE=4 NFE=50
Baseline Teacher (w=1.75) 262.60 221.03 87.56 1.59
FM Only 263.86 220.55 87.93 1.45
CD Only 3.08 2.50 2.54 6.52
FM + CD 1.78 1.41 1.62 1.78
FM + N2N 441.66 297.28 132.60 1.41
FM + CD + N2N (Full DE-CM) 1.70 1.33 1.41 1.26

The ablation study reveals a clear division of labor among the components: CD provides few-step capability (without it, 1-NFE performance degrades to 262+), FM provides stability and multi-step performance (\(FM+CD\) at 50 NFEs yields 1.78, which drops to 1.26 upon adding N2N), and N2N bridges the noise end to the intermediate states (\(FM+CD\) is significantly weaker than the Full configuration at 50 NFEs). None of the three can be omitted.

Hyperparameter ablation (Table 0.D.1): \(\lambda=0.5, \gamma=1\) is optimal (with \(\lambda=0\) yielding 1.73 and \(\lambda=1\) yielding 1.79 for 1-NFE FID). \(\delta=0.1\) (with r-decay) outperforms \(\delta=0\) (no decay, 1.74 vs 1.70). Removing velocity normalization degrades the 1-NFE FID to 1.75. Uniform piecewise timestep sampling outperforms arctan-norm and log-norm. An N2N update frequency of \(\text{freq}=3\) is optimal (\(\text{freq}=1\) is too frequent and causes training instability, while \(\text{freq}=5\) degrades few-step performance).

Key Findings

  • Clear Division of Labor Among Three Losses: CD enables few-step inference, FM stabilizes training and preserves multi-step performance, and N2N eliminates cumulative errors during the first step. Removing any component results in a performance collapse in the corresponding NFE range.
  • Performance Gain at Higher Step Counts: At 50 NFEs, DE-CM achieves an FID of 1.26, outperforming the teacher model (1.59 w/ CFG). This demonstrates that the training process does not disrupt the pre-trained capabilities but rather yields a performance gain.
  • High Efficiency: A mere 16 GPU hours of training achieves a 1-NFE quality superior to sCMs and MeanFlow under equivalent resources (Fig. 8).
  • Implicit Regularization from Velocity Normalization: Beyond preventing oversaturation, bounding the norm of \(\Delta v\) indirectly constrains the numerical range of the JVP vectors, enhancing training stability.

Highlights & Insights

  • Powerful Analysis of Instability in CMs: Decomposing the continuous-time CM objective into fully supervised and self-supervised terms and demonstrating via experiments that instability stems from the JVP fluctuations of the self-supervised term is a clean and solid methodological diagnosis. This can be directly transferred to other self-supervised distillation scenarios.
  • "Selecting Trajectories over Changing Objectives" Mentality: While many works attempt to stabilize CMs by adding constraints or altering architectures, DE-CM does the opposite. By returning to the source of the PF-ODE path and selecting only the most critical subset of trajectories, it achieves \(O(n)\) efficiency and superior stability. This "subtraction-based" thinking is rare but highly effective in the distillation domain.
  • N2N Mapping Enables Truly Flexible Sampling: Traditional CMs are constrained to long jumps with \(\gamma=1\). DE-CM introduces "short-jump" capabilities through N2N mapping. Mapping pure noise to any intermediate state essentially forces the model to internalize the local linear approximation of the entire ODE trajectory, which was previously unattainable for CMs.
  • Rethinking FM as a "Boundary Condition": FM is no longer treated as an auxiliary training task but as a tool to patch the boundary gaps left by the unstable self-supervised term. This perspective of "left boundary regularization" can be generalized to any self-supervised distillation method suffering from instability; introducing fully supervised anchors at \(s=t\) is highly recommended.

Limitations & Future Work

  • JVP Incompatibility with FSDP/Flash Attention: The authors explicitly note that the JVP operator conflicts with the FSDP framework and Flash Attention, leading to high GPU memory consumption and limiting training on larger models. This is a critical engineering bottleneck that requires framework-level improvements or alternative JVP approximations.
  • N2N Update Frequency as an Extra Hyperparameter: The \(\text{freq}=3\) hyperparameter requires tuning. Furthermore, the paper does not delve deeply into the underlying mechanism of why excessively high frequencies cause training instability, while excessively low frequencies hurt few-step performance.
  • An Interesting Contradiction from GAN Integration: Hybrid training with GANs (Table 0.D.2) accelerates convergence during early training but causes an inflection point in high-NFE performance in later stages (with 2/4/50-NFE FIDs increasing after ~160 epochs). This indicates that GANs tend to optimize single-step mappings, which might conflict with the multi-trajectory objectives of DE-CM. This trade-off warrants further investigation.
  • Generalizability of \(r\)-value Decay and Timestep Scheduling: The optimal decay and scheduling strategies might require retuning across different teacher models and resolutions, leaving generalizability to be thoroughly verified.
  • vs. MeanFlow/AYF: Both use the flow mapping \(f_\theta(x_t, t, s)\) to optimize the entire \(\{(t, s)\}\) space via enumeration, leading to \(O(n^2)\) complexity where redundant trajectories slow down convergence. DE-CM selects only three trajectory clusters, achieving \(O(3n)\) efficiency and enhanced stability due to FM providing deterministic anchors. The paper mathematically proves that the objective of DE-CM is equivalent to that of MeanFlow when \(w'(t, s)=1\), with the key difference being full enumeration vs. selective optimization.
  • vs. sCMs: sCMs use TrigFlow parameterization and optimize only a single CM trajectory (\(s=1\), or \(s=T\)), lacking the boundary condition and N2N flexibility. DE-CM extends sCMs by introducing FM boundary regularization and N2N mapping.
  • vs. FACM: FACM decouples the instantaneous velocity objective to improve stability but lacks the N2N mechanism, leading to multi-step error accumulation. DE-CM adopts the decoupling concept and leverages N2N to compensate for the shortfalls in multi-step performance.
  • vs. BOOT: Both optimize path segments starting from pure noise to \(x_t\). However, BOOT is based on bootstrapping of discrete consistency properties (data-free distillation), whereas N2N is a theoretically derived trajectory family based on continuous-time CMs. Their optimization objectives and derivation pathways differ fundamentally.
  • vs. Shortcut Methods: Shortcut relies on FM for anchoring only in low-noise regions near \(d \to 0\), making its stability highly dependent on the model's behavior in those regions. DE-CM extends FM across the entire noise range, acting as a general-purpose left boundary regularizer.

Rating

  • Novelty: โญโญโญโญโ˜† The philosophy of selecting trajectories instead of altering objectives is inspiring, and N2N mapping enables truly flexible sampling in CMs for the first time. However, the selection of the three key trajectory clusters (\(s=1\), \(s=t\), \(r \to 0\)) shares similarities with prior works (CMs, FM, BOOT), leaning more towards a "systematic integration" than a completely pristine concept.
  • Experimental Thoroughness: โญโญโญโญโญ Verified across both C2I and T2I tasks. Tables 1, 2, 4, alongside four ablation tables in the appendix, extensively cover component ablations, hyperparameters, design choices, and GAN hybrid training. It also includes efficiency comparisons (Fig. 5b/8), qualitative comparisons (Fig. 6/7/9), and abundant visualizations in the appendix. The ablation conclusions are clean and convincing, illustrating a clear division of labor.
  • Writing Quality: โญโญโญโญโ˜† The core contributions and motivation analysis are distinct. Mathematical derivations are comprehensive (main text + Appendix 0.A), and the algorithm pseudocode is reproducible. However, the formulas in Sec 4.2 for the N2N derivation are highly dense, and temporary symbol switches (e.g., \(v_\phi\) vs. \(v_\psi\) vs. \(v_\text{tea}\)) might confuse some readers.
  • Value: โญโญโญโญโญ Achieving a 1-NFE FID of 1.70 represents a new SOTA along the CM distillation pathway. The core intuition of the method (performing subtraction on trajectories + utilizing FM as a boundary condition) is concise, transferable, and widely applicable (as validated on both C2I and T2I), carrying substantial practical value for the deployment of few-step generative models.