Diffusion Low-Cost Candidate Exploration (DLCCE): A Unified Framework for Improving Image Generation Quality through Budget-Aware Candidate Exploration, Selection, and Refinement

High-quality image generation with diffusion models can require substantial computational resources, particularly when multiple candidate images must be generated, evaluated, and refined before a satisfactory result is obtained. Conventional generation pipelines often allocate a relatively large proportion of their inference budget to individual generation trajectories, while candidate exploration and subsequent refinement may introduce additional computational costs. This paper proposes Diffusion Low-Cost Candidate Exploration (DLCCE), a unified framework for investigating whether computational resources can be allocated more effectively by initially generating multiple low-cost candidate representations, estimating their potential final quality, and selectively promoting promising candidates to more computationally expensive generation and refinement stages. Unlike a conventional Best-of-N strategy in which all candidates are generated to completion before selection, DLCCE considers candidate exploration as a multi-stage computational allocation problem. Candidate images may initially be generated using reduced resolutions, fewer denoising steps, approximate model evaluations, shortened sampling trajectories, reduced numerical precision, or alternative low-cost generation configurations. Intermediate candidates are then evaluated through one or more inexpensive quality estimators. Based on predicted final quality, uncertainty, diversity, and estimated refinement cost, the framework allocates additional computation to selected candidates. The final output is obtained through selective continuation, resolution enhancement, image-to-image refinement, trajectory optimization, or other quality-improvement operations. The framework formalizes candidate exploration, candidate evaluation, promotion, refinement, and final selection under an explicit end-to-end computational budget. It distinguishes candidate-generation cost from evaluation, ranking, verification, and refinement cost, and considers wall-clock time, model function evaluations, memory consumption, and energy consumption as complementary efficiency measures. Building on the author's previous Lookahead Resolution Allocation (LRA) experiments, this paper extends low-resolution candidate selection into a broader framework that includes resolution, denoising trajectory, noise initialization, candidate diversity, computational fidelity, and adaptive refinement. It also introduces 300 author-proposed DLCCE methods, organized into 20 methodological groups. These formulations are intended to establish a structured collection of proposed techniques that can be investigated individually or combined into larger systems. DLCCE is presented as a conceptual framework and a collection of testable method formulations. It does not claim that all proposed methods have been experimentally validated or that candidate exploration necessarily improves every image-generation task. Its central hypothesis is that, under suitable conditions, inexpensive early exploration and selective computational promotion may provide a more favorable quality–cost trade-off than allocating the full generation budget independently to every candidate.