acceptodds
Under review as a conference paper at ICLR 2027

Compose Once, Sample Many: Amortized Composition of Diffusion Models

Abstract

Composing pretrained diffusion models enables flexible inference-time generation without retraining, but existing composition methods are expensive when many samples are required from the same target distribution: each new sample repeatedly queries all constituent models throughout a long sampling trajectory. We introduce AmCo (Amortized Composition), a two-stage framework that pays this composition cost only for a small set of initial samples and reuses the resulting information for subsequent generation. In Stage 1, AmCo applies an existing composition method to obtain a small set of anchor samples together with their constituent density values. In Stage 2, these anchors are used to approximate the clean prediction of the target composed distribution, while a single constituent model provides a finite-sample correction. Consequently, Stage 2 requires only one constituent-model evaluation per sampling step, and supports substantially shorter deterministic sampling trajectories. Experiments on classifier-free-guided image generation and product-of-experts ligand generation show that AmCo preserves or improves generation quality and diversity while substantially reducing inference cost.

Primary area:
generative models
Keywords:
diffusion models, compositional generation, inference-time sampling, importance sampling

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 2 reviews, above 85% of the others.
Reviewer_VYGZOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the composition of diffusion models, aiming to sample from a composed distribution given access only to the scores of the individual constituent distributions. The authors propose a two-stage method. In the first stage, an existing composition method is used to obtain a set of anchor samples. In the second stage, these anchors are used to construct an empirical score estimate, which is then combined with the score from a single constituent model for sampling. The authors also provide theoretical analyses of the asymptotic estimation error and the optimal correction coefficient. Experiments include a 1D numerical simulation, classifier-free guidance (CFG) for image generation, and ligand generation.

Critical strengths

The paper is well written and easy to follow. The theoretical analysis appears to be sound and provides some insights into the asymptotic behavior of the proposed algorithm.

Critical weaknesses
  1. I am skeptical about whether using only 4–16 anchor samples can provide sufficient information in practice, especially for high-dimensional data such as images, where the empirical score estimated from so few samples could be highly inaccurate. I also noticed that the metrics in Table 8 do not consistently improve as MM increases. In many cases, using more anchors actually leads to worse results. This makes me question whether these anchors really contribute positively to the generation process, or whether similar performance could be achieved using only a single constituent model without any anchors. I would suggest adding at least two simple baselines: (i) directly sampling from the constituent model used in Stage 2, without any anchors; and (ii) more generally, sampling from different constituent models with probabilities determined by the desired composition weights (assuming the weights are nonnegative, so this would not be suitable for extrapolative CFG). I think these trivial baselines are important for demonstrating the actual effectiveness of the anchors. Otherwise, neither my intuition nor some of the experimental results convince me that the anchors are really helpful.

  2. The current experimental section lacks many essential configuration details. For example, which pretrained models and datasets are used for the CFG and ligand generation experiments? How exactly are CLIP Score, ImageReward (IR), and diversity computed? In particular, there are different conventions for computing CLIP Score, so which one is used in this paper? The authors should provide more detailed descriptions or appropriate references for these experimental settings and evaluation metrics.

  3. Another observation is that the adaptive correction coefficient often performs worse than the constant coefficient in Tables 8 and 9. This also makes me question whether the theoretically derived adaptive coefficient is really necessary in practice. I would appreciate a more detailed discussion of this observation and the actual benefits of using the adaptive coefficient over a much simpler constant one.

  4. For the CFG experiments, I think a vanilla CFG baseline is necessary. Specifically, the authors could directly sample using the extrapolated score and report the corresponding evaluation metrics. In addition, I think the image generation experiments should at least include FID results and some visual examples of the generated images.

Questions for authors

For the ligand generation experiments, how is the reference constituent model selected in Stage 2? Is it always the model conditioned on the first protein pocket, as in the other experiments? If so, since the target distribution is symmetric with respect to the two constituent models, I wonder whether choosing a different reference model would significantly affect the results.

8 comments
AuthorsOpenReview ↗

W1(a): Do the Anchors Provide Benefits Beyond Constituent-Only Sampling?

We thank the reviewer for raising this fundamental concern. We agree that estimating a high-dimensional clean-prediction field using only a small number of anchors is challenging.

Importantly, AmCo does not rely on the empirical cache alone. The anchors provide target-specific compositional information, while the reference-model correction incorporates the exact clean prediction from a pretrained constituent model. Nevertheless, we agree that the anchors' practical contribution must be verified experimentally.

We have added the requested constituent-only and constituent-mixture baselines (Tables 8 and 16).

Image generation (Table 8). We directly sample from the conditional constituent used as the Stage-2 reference, without anchors. At the same end-to-end budget of 78 model evaluations per sample, including AmCo's Stage-1 cost:

Guidance wwConstituent-only IRAmCo IR
1.5−0.5900.010
2.0−0.5900.247
2.5−0.5900.273

Thus, sampling from the reference constituent alone does not reproduce AmCo's task-specific performance.

For extrapolative CFG (w>1w>1), the unconditional constituent has exponent 1−w<01-w<0; hence, the composition exponents cannot define nonnegative mixture probabilities. We instead evaluate ordinary extrapolated-score CFG directly (W4).

Ligand generation (Table 16). For product-of-experts composition, we additionally implement the requested constituent-mixture baseline. Each trajectory selects one pocket uniformly at random and samples using only that constituent. We also evaluate the first-pocket constituent and β\beta-scaled ODE controls.

At matched total budgets of 314 evaluations/sample for I=2I=2 and 447 for I=3I=3, AmCo outperforms these controls in mean joint docking score across all 12 evaluated settings.

For example, at β=2.0\beta=2.0, the three-pocket scores are 795.2 (AmCo), 679.8 (first-pocket ODE), and 636.6 (constituent mixture). AmCo also maintains high joint-binding success rates (Table 16).

Although these controls do not sample the exact product-of-experts distribution, they directly demonstrate practical benefits beyond constituent-only generation.

We believe these additions provide direct experimental evidence that the Stage-1 compositional information contributes meaningfully to AmCo's performance.

AuthorsOpenReview ↗

W1(b): Nonmonotonic Cache-Size Behavior and Independent Anchors

We agree that the nonmonotonic dependence on cache size deserves further investigation.

In our main image experiments, the anchors come from a single FKC/RNC particle population with resampling at every step. This can cause genealogical collapse, so increasing the nominal particle count does not necessarily increase the amount of independent information proportionally.

Furthermore, when the total sample count K=M+NK=M+N is fixed, increasing MM changes the proportion of Stage-1 and Stage-2 samples included in the evaluation. Thus, the original trends should not be interpreted as a direct measurement of estimator convergence.

To examine AmCo beyond the particle-collapse regime, we conducted an additional independent-FKC-anchor experiment (Appendix B.2, Table 12, Figure 8).

We independently run FKC 128 times with four particles per run, retaining one output from each. AmCo uses the first MM outputs as anchors and generates N=128−MN=128-M additional samples.

At w=2.0w=2.0, K=128K=128:

MMAmCo IRMatched CFG IRAmCo Div.CFG Div.
8−0.0970.0070.2190.234
160.0300.0310.2260.237
320.0900.0430.2350.239
640.1550.0490.2390.240

AmCo's ImageReward increases as MM increases in this setting. At M=32M=32 and M=64M=64, it outperforms compute-matched CFG, with paired per-class improvements of +0.047 ± 0.021 and +0.106 ± 0.021, while maintaining comparable diversity.

We acknowledge that the improvement with increasing MM also reflects the changing Stage-1/Stage-2 sample proportions. Nevertheless, the matched-cost comparisons show that AmCo retains a practical advantage when using substantially more diverse, independently collected anchors.

We also acknowledge that eight anchors are insufficient to outperform CFG in this setting, and that full independent FKC achieves higher ImageReward (0.268) at considerably greater cost (2,048 evaluations/sample).

These findings clarify that small caches are not universally sufficient; anchor quality, diversity, and cache size all matter. Figure 8 additionally provides qualitative comparisons in this less-collapsed setting.

AuthorsOpenReview ↗

W2: Experimental Configurations and Evaluation Metrics

We agree that the original manuscript lacked sufficient implementation and evaluation details. We have substantially expanded the experimental specifications in Section 4.2 and Appendices B.2–B.3.

Image generation (Appendix B.2).

We use pretrained conditional and unconditional EDM2-XS ImageNet-512 models:

  • Conditional: edm2-img512-xs-2147483-0.045
  • Unconditional: edm2-img512-xs-uncond-2147483-0.045

Both operate on 4×64×644\times64\times64 latents decoded into 512×512512\times512 images using the Stable Diffusion VAE.

We evaluate 100 randomly selected ImageNet classes at w∈{1.5,2.0,2.5}w\in\{1.5,2.0,2.5\}, using T=256T=256 FKC/RNC steps and T′=16T'=16 AmCo steps. Appendix B.2 additionally specifies the noise schedule, finite-noise endpoint, resampling settings, and correction hyperparameters.

Metric definitions.

We now explicitly specify the CLIPScore convention of Hessel et al. (2021):

CLIPScore=250max⁡(cos⁡(EI,ET),0),\mathrm{CLIPScore} =250\max(\cos(E_I,E_T),0),

using OpenAI CLIP ViT-B/32, with the text prompt a photo of a <class name> and normalized image/text embeddings.

ImageReward is evaluated using the official ImageReward-v1.0 model (Xu et al., 2023). Diversity is defined as one minus the mean pairwise cosine similarity of CLIP ViT-B/32 image embeddings within each class.

We also specify the conventions for FID, FDDINOv2, CMMD, Inception Score, and ConvNeXt classification accuracy (Tables 5–8).

Ligand generation (Appendix B.3).

We use the target-aware molecular diffusion architecture of Guan et al. (2023). The revised appendix specifies the evaluated protein-pocket tuples, two- and three-pocket compositions, molecular size, sampling parameters, and categorical atom-type handling.

Docking is evaluated using Uni-Dock. We define Jdock=∏i∣Pi∣J_{\mathrm{dock}}=\prod_i|P_i| over molecules binding all target pockets, and molecular diversity using Tanimoto similarity. Table 16 additionally reports validity and joint-binding success rates.

These revisions substantially improve the clarity and reproducibility of our experimental evaluation.

Reviewer_CLQbOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes AmCo, a two-stage framework for reducing the cost of repeatedly sampling from a composition of multiple diffusion models. First, a composition sampler generates a small set of anchor samples and caches their constituent density estimates. Then, the anchors are used to approximate the clean prediction of the composed distribution through importance weighting. A single constituent diffusion model is then used to correct the finite-sample estimation error, with a jackknife approach.

The author provides an asymptotic analysis of the corrected estimator and evaluates the method on a controlled 1D composition task, classifier-free-guided image generation, and ligand generation. The results show substantial reductions in inference cost while maintaining competitive generation quality.

Critical strengths

1. A simple and effective approach to amortizing diffusion composition. The core idea of reusing a small set of composed samples to avoid repeated evaluations of all constituent models is strongly motivated. The reference-model correction provides a principled way to address the finite-sample limitations of the cached estimator, and the theoretical analysis supports explain its reduction of trace ofthe covariance. Additionally, AmCo requires no additional model training and this point is also appealing.

2. Strong computational improvements supported by multiple experiments. The paper evaluates the proposed correction using clean-prediction MSE and Wasserstein distance against an exact target in the 1D setting. Moreover, the image and ligand experiments further demonstrate substantial computational savings while maintaining or improving the reported quality and diversity metrics. In particular, the image experiments report a 4.0x reduction in total FLOPs even after accounting for CREPE initialization and burn-in, which makes the practical benefit more convincing.

Critical weaknesses

W1. Limited theoretical characterization of the actual sampling process. Theorem 1 establishes the asymptotic behavior of the corrected clean-prediction estimator for a fixed diffusion time and query point, assuming i.i.d. anchor samples. The 1D experiments also provide empirical evidence of endpoint distribution accuracy through Wasserstein distance. However, during Stage 2, the query states are generated recursively using the same finite cache, making them dependent on the anchor samples. The current analysis does not characterize how approximation errors accumulate along this trajectory or affect the final generated distribution. This leaves a gap between the theoretical results and the sampling procedure used in practice.

W2. Limited comparisons under matched computational budgets. The paper provides detailed FLOP comparisons against CREPE and FKC, including the initialization and burn-in costs of CREPE. The appendix also evaluates different cache sizes and correction strategies. These results establish substantial savings relative to the original composition procedures. However, AmCo also uses considerably fewer sampling steps, and the paper does not compare its generation quality against accelerated alternatives under similar computational budgets. Such comparisons would help determine whether the proposed amortization provides a better quality–cost trade-off than other approaches to reducing sampling cost.

Questions for authors

Q1. How sensitive is the method to the choice of reference model? Tables 8 and 9 examine the effect of different correction coefficients, but the reference model itself is not varied. In the CFG experiments, the conditional model is used as the reference. Would choosing the unconditional model lead to similar results? Since the correction depends on the relationship between the reference and target estimation errors, it would be helpful to understand how much performance depends on this choice.

Q2. How sensitive is the correction to errors in the cached density estimates? The importance weights in Equation (16) depend on constituent density values estimated during Stage 1. While the controlled 1D experiments validate the estimator using analytically specified distributions, the effect of density-estimation error in practical diffusion models remains unclear. Could the authors evaluate how controlled perturbations to the cached log-density estimates affect the corrected clean prediction and generation quality?

6 comments
AuthorsOpenReview ↗

W1: Theoretical Analysis of Cache-Dependent Sampling Trajectories

We thank the reviewer for identifying the gap between pointwise clean-prediction consistency and the behavior of the recursively generated Stage-2 sampler.

We agree that Theorem 1 alone does not establish convergence of the final sampling distribution, since the Stage-2 query states depend on the same finite Stage-1 cache.

To address this issue, we added Lemma 6 (Appendix A.3, p. 21), establishing a cache-conditional trajectory stability bound.

The analysis considers the actual DDIM-style Stage-2 update (Eq. 21) and compares AmCo with an oracle discretized sampler using the exact target clean prediction, under the same initialization and time discretization.

Under a Lipschitz assumption on the oracle clean-prediction field, we establish

∥X0−Y0∥≤∑j=1T′Ajej,\|X_0-Y_0\|\leq\sum_{j=1}^{T'}A_j e_j,

where eje_j is the clean-prediction error evaluated at the actual AmCo query state, and AjA_j describes how this error propagates through subsequent sampling steps.

This yields the endpoint Wasserstein bound

\mathcal L(Y_0)\bigr) \leq \sum_{j=1}^{T'} A_j \bigl(\mathbb E[e_j^2\mid\mathcal C]\bigr)^{1/2}.$$ **Crucially, the proof conditions on the cache and does not require independence between the cache and the recursively generated query states.** It therefore directly addresses the dependence highlighted by the reviewer. We also clarify the scope of this result. The lemma characterizes how prediction errors accumulate along the trajectory, but **pointwise convergence alone does not imply an $O(M^{-1/2})$ endpoint convergence rate** without additional trajectory-level error control. Discretization and finite-noise endpoint errors are separate approximation sources. Finally, our controlled 1D experiments directly measure endpoint Wasserstein distance against the exact composed target (Section 4.1, Figure 2, Table 4), complementing the theoretical stability analysis. We believe this addition substantially strengthens the connection between our estimator-level theory and the actual sampling procedure.
AuthorsOpenReview ↗

W2: Compute-Matched Comparisons and Quality–Cost Trade-offs

We agree that lower computational cost relative to a long composition trajectory does not by itself demonstrate an advantage over accelerated sampling alternatives.

We have therefore added exactly compute-matched baselines and quality–total-cost comparisons, including all Stage-1 cache construction costs.

Image generation (Table 8, p. 30).

In our main setting, M=8,N=56,T=256,T′=16M=8,N=56,T=256,T'=16, AmCo requires

8(256)(2)+56(16)=4,9928(256)(2)+56(16)=4,992

model evaluations for 64 images, corresponding to 78 evaluations per sample, including Stage 1.

We compare against 39-step ordinary CFG ODE, 78-step conditional-only ODE, and 39-step FKC/RNC, all under the same total evaluation budget.

The mean ImageReward results are:

Guidance wwCFG ODEAmCo
1.5−0.2380.010
2.0−0.0490.247
2.50.0490.273

The conditional-only ODE achieves −0.590 across these settings.

Thus, AmCo provides a task-quality advantage over both ordinary CFG and reference-only sampling under matched total computation.

We also report CMMD, a CLIP-based kernel metric proposed to address limitations of FID (Jayasumana et al., CVPR 2024). AmCo achieves lower CMMD than matched-cost CFG across all three guidance strengths: (0.650,0.514,0.481)(0.650,0.514,0.481) versus (0.672,0.550,0.495)(0.672,0.550,0.495).

We acknowledge that CFG achieves better FID and generally greater diversity. Accordingly, these results support a favorable quality–cost trade-off rather than uniform superiority across all metrics.

Sampling-budget sweeps (Table 9, Figure 5).

We additionally evaluate FKC/RNC over T∈T\in{32,64,128,256}, across three guidance strengths and 100 ImageNet classes.

At the same TT, AmCo achieves higher ImageReward in 22 of 24 configurations. At T=256T=256, AmCo requires 4,992 evaluations compared with 32,768 for full composition sampling, a 6.6× reduction. Figure 5 plots the resulting quality–total-cost trade-offs.

Ligand generation (Table 16, p. 40).

We also introduce matched-budget single-pocket and constituent-mixture ODE baselines, accounting for both Stage-1 costs and categorical atom-type refreshes.

At 314 evaluations/sample for two-pocket composition and 447 for three-pocket composition, AmCo achieves higher mean joint docking scores than these controls across all 12 evaluated configurations, with joint-binding success rates of 97.9–100%.

These single-constituent controls do not target the exact product-of-experts distribution, but provide relevant practical cost–quality comparisons.

Together, these experiments demonstrate that AmCo's computational benefits are not solely attributable to reducing Stage-2 sampling steps.

AuthorsOpenReview ↗

Q1: Sensitivity to the Choice of Reference Model

We thank the reviewer for raising this question. Since the reference correction exploits correlated estimation errors between the target and reference predictions, its effectiveness can depend on the selected constituent.

We have added reference-model swap experiments for both image and ligand generation.

Image generation (Tables 5–6, pp. 27–29).

We compare conditional and unconditional EDM2-XS reference models while holding the Stage-1 cache and Stage-2 initial noise fixed.

For w=2.0w=2.0, (M,N)=(8,56)(M,N)=(8,56):

Stage-1 samplerConditional ref. IRUnconditional ref. IR
FKC0.2470.236
RNC−0.021−0.108

The two references give comparable ImageReward under FKC, whereas the conditional reference performs noticeably better under RNC.

These results show that the reference correction can operate with either constituent, but its practical effectiveness is not invariant to reference selection.

Ligand generation (Tables 13–14, pp. 37–38).

We also vary the reference protein pocket for both two- and three-target compositions.

For example, with RNC, I=3I=3, β=1.5\beta=1.5, and (M,N)=(4,28)(M,N)=(4,28), the joint docking scores are:

Reference pocketJdockJ_{\mathrm{dock}} ↑
Pocket 1935.6
Pocket 2891.7
Pocket 3958.8

Thus, the first pocket is not always the optimal reference.

Although the ideal product-of-experts target is symmetric with respect to its constituents, the finite-cache estimator and reference correction need not be. Different reference models may exhibit different error correlations with the target estimator.

We have included these experiments to clarify both the flexibility and the practical sensitivity of reference selection.

Comments on the paper

7 comments
AuthorsOpenReview ↗

Summary of Revisions

We sincerely thank all reviewers for their constructive feedback. In response, we have revised the manuscript, including a redesign of the image-generation experiments, additional theoretical analysis, and extensive new experiments and ablations.

Major revisions are highlighted in blue in the updated manuscript. We summarize the changes below.

1. Redesigned image-generation experiments (Nqjh W1/W2; CLQb W2; VYGZ W1/W4)

We replaced CREPE with Feynman–Kac Corrector (FKC) and Radon–Nikodym Corrector (RNC) as the base composition methods for image generation.

To address concerns about computational fairness and generation quality, we added:

  • Compute-matched baselines, including vanilla CFG and constituent-only ODE sampling (Table 8).
  • Sampling-budget ablations over T∈{32,64,128,256}T\in\{32,64,128,256\} and quality–cost curves (Table 9, Figure 5).
  • Additional image metrics, including FID, FDDINOv2, CMMD, Inception Score, and classification accuracy (Tables 5–8).
  • Qualitative comparisons showing Stage-1 anchors, Stage-2 outputs, and baseline samples (Figures 4 and 8).

All compute-matched comparisons account for Stage-1 cache construction costs.

2. Strengthened theoretical analysis (CLQb W1)

We added Lemma 6 (Appendix A.3), establishing a cache-conditional stability bound for the recursive Stage-2 sampler.

The result relates endpoint Wasserstein error to accumulated clean-prediction errors along the generated trajectory, without requiring independence between the cache and query states.

We also clarify that pointwise convergence alone does not establish an endpoint convergence rate.

3. Expanded robustness and ablation studies (CLQb Q1/Q2; VYGZ W1/W3/Q)

We added:

  • Density-tracking perturbation experiments, evaluating their effects on corrected clean predictions and generation quality (Tables 10–11, Figures 6–7).
  • Correction-coefficient ablations, comparing no correction, constant correction, and adaptive jackknife correction, including nearest-anchor distances (Tables 17–18).
  • Reference-model sensitivity analyses in both image and ligand generation (Tables 5–6, 13–14).
  • Compute-matched constituent-only and constituent-mixture baselines for ligand generation (Table 16).

Additional experiments

Beyond the specific reviewer requests, we also conducted two additional experiments to examine the generality and robustness of AmCo:

  • Independent FKC anchors: We collected anchors from independently initialized FKC runs to evaluate AmCo without severe within-population particle collapse (Table 12, Figure 8).
  • Three-model composition: We extended ligand generation from two to three constituent models (I=3I=3) to investigate whether AmCo remains effective as the number of composed models increases (Tables 14–16).

Closing

The revised manuscript provides stronger evidence regarding computational efficiency, anchor utility, generation quality, reference sensitivity, and estimator stability.

Detailed responses to individual concerns are provided in the corresponding reviewer discussion threads. We thank the reviewers again for their constructive suggestions.

AuthorsOpenReview ↗

W1/Q1: Original CREPE Configuration and Revised Samplers

We thank the reviewer for questioning the original 2048-step CREPE configuration.

Our original T=2048T=2048 was motivated by numerical stability of the Radon–Nikodym Estimator (RNE) used to track constituent log-densities. It was a conservative implementation choice, not a requirement of CREPE or an optimized sampling budget. We agree that the absence of shorter-sampler controls made the original efficiency comparison insufficient.

We have therefore replaced CREPE with Feynman–Kac Corrector (FKC) and Radon–Nikodym Corrector (RNC) in the image experiments (Sec. 4.2; App. B.2), for three reasons.

First, CREPE incurs a substantial burn-in overhead relative to AmCo's small Stage-1 cache. Our original implementation generated 168 sequential outputs, discarded 40 burn-in samples, and retained 128. Although we accounted for initialization and burn-in costs, this fixed overhead is relatively expensive when only a small number of anchors are needed. In contrast, FKC/RNC generate particle populations without a separate sequential MCMC burn-in phase.

Second, FKC/RNC enable systematic and fair quality–cost comparisons. We can directly compare Stage-1 cache construction and full composition sampling using the same sampler, diffusion schedule, pretrained models, and evaluation protocol, while varying the computational budget.

Third, the statistical properties of SMC particles are more directly connected to our theoretical setting. CREPE produces sequentially correlated MCMC samples, whereas our theoretical analysis assumes i.i.d. anchors. Although finite-population FKC/RNC particles are not strictly independent, standard SMC propagation-of-chaos results guarantee asymptotic independence for any fixed number of particles as the population size increases, under appropriate regularity assumptions. We additionally investigate finite-population dependence using independently collected FKC anchors (Table 12).

The revised efficiency claims are based on the new FKC/RNC experiments, including sampling-budget sweeps and compute-matched comparisons, rather than the original 2048-step CREPE comparison.

AuthorsOpenReview ↗

W1/Q1: Quality–Cost Trade-off and Compute-Matched Baselines

We agree that comparing only against long composition trajectories does not establish a practical computational advantage.

We have added sampling-budget sweeps, quality–total-cost curves, and exactly compute-matched baselines, accounting for Stage-1 cache construction.

1. Sampling-budget sweep (Table 9, Fig. 5).

We evaluate FKC and RNC over T∈T\in{32,64,128,256}, three guidance scales, and 100 ImageNet classes.

At T=256T=256, generating 64 samples costs:

  • AmCo: 8(256)(2)+56(16)=4,9928(256)(2)+56(16)=4,992 evaluations.
  • Full FKC/RNC: 64(256)(2)=32,76864(256)(2)=32,768 evaluations.

Thus, AmCo requires 6.6× fewer evaluations, including Stage 1, while achieving higher mean ImageReward in 22 of 24 matched-TT settings. Figure 5 plots quality against total computational cost.

2. Exactly matched-budget comparison (Table 8).

AmCo requires 78 evaluations per image on average, including cache construction. We compare against 39-step vanilla CFG ODE, 78-step conditional-only ODE, and 39-step FKC/RNC.

At guidance scales w=(1.5,2.0,2.5)w=(1.5,2.0,2.5):

  • AmCo IR: (0.010,0.247,0.273)
  • CFG ODE IR: (-0.238,-0.049,0.049)
  • Conditional-only IR: -0.590

AmCo achieves higher ImageReward than these baselines across all three guidance scales.

We also include inexpensive 16-evaluation controls. Since revised Stage 2 uses T′=16T'=16, rather than the original 64, we compare against both 16-step marginal-cost and 78-step end-to-end matched-cost conditional-only sampling.

These results provide direct evidence that AmCo's advantage is not merely due to shortening its sampling trajectory.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.