OpenNovelty.org

HiGS: History-Guided Sampling for Plug-and-Play Enhancement of Diffusion Models

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 6.0Download Report PDF
diffusion modelssamplingclassifier-free guidance
Abstract:

While diffusion models have made remarkable progress in image generation, their outputs can still appear unrealistic and lack fine details, especially when using fewer number of neural function evaluations (NFEs) or lower guidance scales. To address this issue, we propose a novel momentum-based sampling technique, termed history-guided sampling (HiGS), which enhances quality and efficiency of diffusion sampling by integrating recent model predictions into each inference step. Specifically, HiGS leverages the difference between the current prediction and a weighted average of past predictions to steer the sampling process toward more realistic outputs with better details and structure. Our approach introduces practically no additional computation and integrates seamlessly into existing diffusion frameworks, requiring neither extra training nor fine-tuning. Extensive experiments show that HiGS consistently improves image quality across diverse models and architectures and under varying sampling budgets and guidance scales. Moreover, using a pretrained SiT model, HiGS achieves a new state-of-the-art FID of 1.61 for unguided ImageNet generation at 256×\times256 with only 30 sampling steps (instead of the standard 250). We thus present HiGS as a plug-and-play enhancement to standard diffusion sampling that enables faster generation with higher fidelity.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper proposes History-Guided Sampling (HiGS), a momentum-based technique that integrates weighted averages of past predictions to enhance diffusion model sampling quality and efficiency. It resides in the 'Momentum-Based History Integration' leaf under 'General-Purpose Sampling Enhancement', sharing this leaf with only one sibling paper (Adaptive Momentum Sampler). This places the work in a relatively sparse research direction within the broader taxonomy of 23 papers across the field, suggesting the specific approach of momentum-driven history integration for general diffusion sampling remains underexplored compared to task-specific forecasting or spatiotemporal methods.

The taxonomy reveals neighboring leaves focused on extrapolation-based acceleration, computational reuse across prompts, and knowledge distillation, all within the same 'General-Purpose Sampling Enhancement' branch. These directions share the goal of improving diffusion efficiency but diverge in mechanism: extrapolation methods predict future states rather than guide via momentum, while computational reuse exploits prompt similarity rather than prediction history. The broader field includes substantial activity in time series forecasting and spatiotemporal generation, where history mechanisms serve domain-specific constraints (e.g., temporal coherence, trajectory consistency) rather than general sampling quality. HiGS's position suggests it bridges general-purpose efficiency with momentum principles, distinct from both task-specific and prediction-based acceleration approaches.

Among 30 candidates examined, the contribution-level analysis shows varying degrees of prior overlap. The core HiGS method examined 10 candidates with 3 appearing to provide refutable prior work, indicating some existing momentum or history-based sampling techniques in the limited search scope. The plug-and-play enhancement claim examined 10 candidates with 6 potentially refutable, suggesting training-free diffusion improvements are more established in the examined literature. The state-of-the-art FID result examined 10 candidates with only 2 refutable, implying the specific performance benchmark may be less contested among the top-30 semantic matches. These statistics reflect a focused search rather than exhaustive coverage, leaving open the possibility of additional relevant work beyond the examined scope.

Based on the limited search of 30 semantically similar papers, HiGS appears to occupy a moderately explored niche within momentum-based diffusion sampling. The sparse population of its taxonomy leaf and the contribution-level statistics suggest the specific combination of history-weighted momentum and training-free integration may offer incremental novelty, though the analysis cannot rule out additional prior work outside the top-K semantic matches or in adjacent research communities not fully captured by the taxonomy structure.

Taxonomy

Core-task Taxonomy Papers
23
3
Claimed Contributions
30
Contribution Candidate Papers Compared
11
Refutable Paper

Research Landscape Overview

Core task: Enhancing diffusion model sampling quality and efficiency using prediction history. The field has organized itself around four main branches. General-Purpose Sampling Enhancement encompasses methods that improve diffusion sampling across diverse data types, often by incorporating momentum-based or adaptive strategies that leverage past predictions to accelerate convergence or refine outputs. Time Series and Sequential Forecasting focuses on applying diffusion models to temporal data, where historical context naturally informs future predictions. Spatiotemporal Prediction and Generation addresses problems involving both spatial and temporal dimensions, such as video synthesis or trajectory forecasting, where prediction history can guide coherent evolution across frames. Domain-Specific Applications tailors these history-aware techniques to specialized tasks like motion generation, image quality assessment, or event modeling, demonstrating how prediction reuse adapts to particular constraints and objectives. Within General-Purpose Sampling Enhancement, a particularly active line explores momentum-based integration, where methods like Adaptive Momentum Sampler[14] and HiGS[0] incorporate gradient or prediction history to stabilize and speed up the reverse diffusion process. HiGS[0] sits squarely in this momentum-driven cluster, emphasizing how accumulated prediction signals can guide sampling steps more efficiently than memoryless approaches. This contrasts with works in Time Series and Sequential Forecasting, such as Sequential Posterior Sampling[3] or Auto-regressive Moving Diffusion[4], which exploit temporal dependencies inherent to sequential data rather than general momentum schemes. Meanwhile, Spatiotemporal Prediction branches like History-Guided Video[21] or SocialTraj[20] face the challenge of maintaining consistency across both space and time, often requiring more complex history mechanisms than purely temporal or momentum-based methods. The central trade-off across these directions is between the generality of history integration and the specificity needed for coherent long-range or domain-constrained generation.

Claimed Contributions

History-Guided Sampling (HiGS) method

The authors introduce HiGS, a training-free sampling method that leverages a weighted average of past model predictions to guide the diffusion sampling process. This momentum-based approach improves image quality, sharpness, and structural coherence, especially under low NFE or low CFG regimes.

10 retrieved papers
Can Refute
Plug-and-play enhancement requiring no training or fine-tuning

HiGS is designed as a plug-and-play modification that adds negligible computational overhead and can be directly applied to pretrained diffusion models without retraining or architectural changes.

10 retrieved papers
Can Refute
State-of-the-art FID for unguided ImageNet generation

By applying HiGS to a pretrained SiT model, the authors achieve a state-of-the-art FID score of 1.61 on ImageNet 256×256 without classifier-free guidance, using only 30 steps compared to the baseline's 250 steps.

10 retrieved papers
Can Refute

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

History-Guided Sampling (HiGS) method

The authors introduce HiGS, a training-free sampling method that leverages a weighted average of past model predictions to guide the diffusion sampling process. This momentum-based approach improves image quality, sharpness, and structural coherence, especially under low NFE or low CFG regimes.

Contribution

Plug-and-play enhancement requiring no training or fine-tuning

HiGS is designed as a plug-and-play modification that adds negligible computational overhead and can be directly applied to pretrained diffusion models without retraining or architectural changes.

Contribution

State-of-the-art FID for unguided ImageNet generation

By applying HiGS to a pretrained SiT model, the authors achieve a state-of-the-art FID score of 1.61 on ImageNet 256×256 without classifier-free guidance, using only 30 steps compared to the baseline's 250 steps.

Table of Contents