Convergence of Consistency Model with Multistep Sampling under General Data Assumptions

Authors: Yiding Chen, Yiyi Zhang, Owen Oertell, Wen Sun

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Motivations: Consistency model has already demonstrated its power on large-scale image generation tasks (Luo et al., 2023; Song et al., 2023; Song & Dhariwal, 2024). To verify our theoretical findings, we focus on a toy example that is easier to interpret. We first refine our upper bound in Theorem 2, where we relax our result for a cleaner presentation. We make adjustment to (15) and get: supx,y supp(Pdata) x y 2 α2 t1 2σ2 t1 Ex Pdata h x 2 2 i + PN j=2 α2 tj 4σ2 tj t2 j 1 ϵ2 cm τ 2 1/4 + t N ϵcm τ . (26) Simulation setting: We consider OU process as the forward process, which is our setup in Case study 1. For simplicity, we consider a Bernoulli data distribution: Prx Pdata[x = 0] = Prx Pdata[x = 100] = 0.5. This data distribution ensures a close-form for the ground truth consistency function: f (x, t) := ( 0 if x < 50 exp( t) 100 o.w. . We construct a perturbed ˆf( , ) accordingly: ˆf(x, t) := ( 0 if x < at 100 o.w. , where the sequence at satisfies: Prx Pt[x < at] = 0.5 + 0.0001t2, t. This choice of ˆf( , ) makes sure: ˆf(x, t) f (x, t) 2 This means ˆf( , ) satisfies the first statement of Lemma 2 with ϵ2 cm τ 2 = 1. We simulate three instantiations of {ti}N i=1 defined in (5), i.e. the sequence of time steps for our multi-step sampling defined in (5): our schedule: the two-step schedule suggested by Case study 1. We also calculate the upper bound in (26) for comparison; baseline 1: design the sequence of sampling time steps by evenly dividing an interval; baseline 2: start with some T and reduce it by half every step until reaching a small value. In Figure 2, we plot the W2 error in multi-step sampling. We present the revolution of W2 error in a sampling time schedule on a single curve. Specifically, we plot each curve by: ti, W2( ˆP (i) 0 , Pdata) i = 1, ..., N. Because the sampling time step ti decreases in the multi-step sampling by definition. We reverse the x-axis of the plot for presentation purposes. Figure 2: W2 error in multi-step sampling. Observations: This simulation result demonstrates that: Our upper bound is a reasonable characterization of the performance for the designed sampling time schedule. The two-step sampling time schedule suggested by Case study 1 achieves comparable performance to the best result in the baseline methods but with a much smaller number of function evaluations; Running too many sampling time steps may degrade the sampling quality. The error increases for both baseline methods in the last few sampling steps.
Researcher Affiliation Academia 1Cornell University. Correspondence to: Yiding Chen <EMAIL>.
Pseudocode Yes For completeness, we summarize this process in Algorithm 1 in Section B. ... Algorithm 1 Multistep Consistency Sampling
Open Source Code No The paper does not provide any statement or link regarding the availability of source code for the methodology described.
Open Datasets No Simulation setting: We consider OU process as the forward process, which is our setup in Case study 1. For simplicity, we consider a Bernoulli data distribution: Prx Pdata[x = 0] = Prx Pdata[x = 100] = 0.5. This data distribution ensures a close-form for the ground truth consistency function: f (x, t) := ( 0 if x < 50 exp( t) 100 o.w. . We construct a perturbed ˆf( , ) accordingly: ˆf(x, t) := ( 0 if x < at 100 o.w. , where the sequence at satisfies: Prx Pt[x < at] = 0.5 + 0.0001t2, t. This choice of ˆf( , ) makes sure: ˆf(x, t) f (x, t) 2 This means ˆf( , ) satisfies the first statement of Lemma 2 with ϵ2 cm τ 2 = 1. The paper uses a custom-defined Bernoulli data distribution for its simulation and does not refer to any publicly available datasets with access information.
Dataset Splits No The paper uses a custom-defined Bernoulli data distribution for its simulation. It does not mention any training, validation, or test splits, as is typical for purely theoretical or toy example-based simulations.
Hardware Specification No The paper mentions "Our simulation in Appendix I" but does not provide any specific details about the hardware used to run this simulation or any other experiments.
Software Dependencies No The paper does not specify any software dependencies with version numbers.
Experiment Setup Yes Simulation setting: We consider OU process as the forward process, which is our setup in Case study 1. For simplicity, we consider a Bernoulli data distribution: Prx Pdata[x = 0] = Prx Pdata[x = 100] = 0.5. This data distribution ensures a close-form for the ground truth consistency function: f (x, t) := ( 0 if x < 50 exp( t) 100 o.w. . We construct a perturbed ˆf( , ) accordingly: ˆf(x, t) := ( 0 if x < at 100 o.w. , where the sequence at satisfies: Prx Pt[x < at] = 0.5 + 0.0001t2, t. This choice of ˆf( , ) makes sure: ˆf(x, t) f (x, t) 2 This means ˆf( , ) satisfies the first statement of Lemma 2 with ϵ2 cm τ 2 = 1. We simulate three instantiations of {ti}N i=1 defined in (5), i.e. the sequence of time steps for our multi-step sampling defined in (5): our schedule: the two-step schedule suggested by Case study 1. We also calculate the upper bound in (26) for comparison; baseline 1: design the sequence of sampling time steps by evenly dividing an interval; baseline 2: start with some T and reduce it by half every step until reaching a small value.