Guided Trajectory Optimization with Sparse Scaling for Test-Time Diffusion

2026-05-22 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, extended

Summary

The Reward-guided Trajectory Scaling (RTS) method significantly enhances diffusion model generation performance by addressing limitations in existing Test-Time Scaling (TTS) solutions. RTS introduces a novel approach that combines a reward-guided noise optimization strategy with a sparse test-time scaling framework. The optimization actively directs the search towards high-reward noise regions, while the sparse framework, utilizing PCA-driven curvature analysis, prioritizes key intermediate denoising steps to efficiently compress the search space. Experiments demonstrate RTS outperforms baselines, achieving a 15.6% improvement in GenEval Score and a 60.4% enhancement in ImageReward score. Tested on Stable Diffusion v3, FLUX-1.dev, and Qwen-Image across GenEval and T2I-CompBench benchmarks, RTS achieves leading performance, boosting ImageReward by approximately 66.0% on SD v3 and 60.5% on FLUX-1.dev.

Key takeaway

For Machine Learning Engineers aiming to enhance diffusion model performance without costly retraining, the Reward-guided Trajectory Scaling (RTS) method offers a compelling solution. You should explore integrating sparse optimization, using PCA to identify critical denoising steps, and a reward-guided coarse-to-fine search for initial and intermediate noise. This approach can yield substantial gains in image quality and semantic alignment, as demonstrated by its 60.4% ImageReward boost, making your inference more efficient and effective.

Key insights

Reward-guided Trajectory Scaling (RTS) optimizes diffusion model noise paths via sparse, PCA-driven key-step selection and a coarse-to-fine search.

Principles

• Jointly optimize initial and intermediate noises for full generative potential.
• Sparse optimization via PCA curvature analysis identifies critical denoising steps.
• Coarse-to-fine search with surrogate gradients efficiently explores high-reward noise regions.

Method

RTS employs PCA-driven curvature analysis to select sparse key denoising steps. It then applies a reward-guided coarse-to-fine alternating search, estimating surrogate gradients from neighbor samples to optimize initial and intermediate noises.

In practice

• Apply PCA to identify critical denoising trajectory inflection points.
• Implement coarse-to-fine search with spherical interpolation for noise exploration.
• Estimate surrogate gradients from neighbor sample rewards for guidance.

Topics

• Test-Time Scaling
• Diffusion Models
• Image Generation
• Noise Optimization
• PCA
• Reward-Guided Learning

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• Simple Approximation and Derivative Free Inference-Time Scaling for Diffusion Models via Sequential Monte Carlo on Path Measures — URGE improves diffusion model sample quality via derivative-free pathwise importance reweighting and resampling.
• Colored Noise Diffusion Sampling — Diffusion model SDE inference can be optimized by dynamically allocating noise energy based on frequency resolution.
• Initialization-Aware Score-Based Diffusion Sampling — Optimizing the backward process initialization in SGMs significantly reduces computational cost while maintaining generative quality.
• Wasserstein Convergence of ODE-Based Samplers in Decentralized Diffusion Model via Velocity Field Decomposition — The first W₂ convergence guarantee for decentralized diffusion models with ODE-based sampling is established, showing ℮(N⁻¹/²+ε) rate.
• ResilPhase: Plug-and-Play Phase Mapping and Noise-Resilient Macro-Trajectory Extrapolation for Diffusion Acceleration — ResilPhase accelerates diffusion models by reformulating inference as stable ODE-space macro-trajectory extrapolation, using derivative-free methods and phase mapping.
• Diffusion Alignment Beyond KL: Variance Minimisation as Effective Policy Optimiser — VMPO minimizes log importance weight variance for diffusion alignment, unifying existing methods and suggesting new designs.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.