Inference-Time Scaling of Diffusion Models via Progressive Seed Pruning

2026-07-23 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision & Pattern Recognition · Depth: Expert, quick

Summary

Progressive Seed Pruning (PSP) addresses the underdeveloped inference-time scaling for diffusion and flow-matching models, which currently struggle with final quality sensitivity to initial noise seeds. Unlike traditional methods that maintain a constant memory footprint and spend extra compute on seed search, PSP relaxes this constraint by front-loading exploration. It evaluates numerous seeds early in the process and aggressively prunes less promising candidates, thereby utilizing a fixed compute budget more effectively. PSP operates by scoring intermediate denoised estimates and progressively narrowing the candidate set, ensuring only high-potential trajectories are fully denoised while keeping the total number of model evaluations constant. This technique consistently improves reward-guided selection, yielding higher GenEval scores and superior human evaluation for prompt-alignment compared to best-of-N, importance-sampling, and tree-search baselines, all at matched compute.

Key takeaway

For Machine Learning Engineers optimizing diffusion or flow-matching model inference, Progressive Seed Pruning (PSP) offers a significant efficiency gain. If you are struggling with compute budgets or the quality sensitivity to initial noise seeds, you should consider implementing PSP. It allows you to achieve superior prompt-alignment and automated evaluation scores by intelligently pruning less promising trajectories early, making your fixed compute budget more effective than traditional best-of-N or tree-search methods.

Key insights

Optimizing diffusion model inference involves early, aggressive pruning of noise seeds to maximize compute efficiency.

Principles

• Relaxing memory constraints enables new scaling axes.
• Front-loading exploration improves compute budget use.
• Intermediate scoring guides trajectory selection.

Method

PSP scores intermediate denoised estimates, progressively narrowing the candidate set to fully denoise only promising trajectories, while maintaining a fixed total number of model evaluations.

In practice

• Apply PSP to improve reward-guided selection.
• Use PSP for higher GenEval scores in image generation.

Topics

• Diffusion Models
• Flow-Matching Models
• Inference Optimization
• Seed Pruning
• Image Generation
• Compute Efficiency

Code references

• rogerioagjr/psp

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

Related on AIssential

• Importance-Aware OBS Pruning for Diffusion Models — Content-aware pruning, using spatial importance maps, enables perceptually faithful compression of diffusion models at high ratios.
• PPDM: Pixel Puzzling Diffusion Model for Speed and Memory Efficient Volumetric Medical Image Translation — PPDM efficiently translates 3D medical images by trading spatial resolution for channel depth and focusing on task-relevant residuals.
• 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.
• Finding Sparse Subnetworks in One Training Cycle via Progressive Magnitude-Based Pruning — Progressive magnitude-based pruning enables effective neural network sparsification within a single training cycle.
• OrderDP: A Theoretically Guaranteed Lossless Dynamic Data Pruning Framework — OrderDP provides theoretically guaranteed, unbiased, and near-lossless data pruning for accelerated model training.
• S4oP: Operator-level Pruning of Structured State Space Models for Resource-Constrained Devices — Operator-level pruning significantly reduces SSM inference costs while maintaining performance on constrained devices.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.