Lookahead Resolution Allocation (LRA): Third Experimental Report — A 675-Condition Evaluation of Low-Resolution Candidate Ranking, High-Resolution Refinement, and Direct Best-of-N Generation

This third experimental report investigates Lookahead Resolution Allocation (LRA) through a larger controlled evaluation involving 675 experimental conditions. The study examines whether low-resolution candidate exploration followed by selective high-resolution refinement can provide a useful quality–computation trade-off compared with direct high-resolution generation and direct Best-of-N selection. The experiment evaluates nine methods across 15 text prompts and five random seeds. The methods include single-pass direct generation at 768 × 768, direct Best-of-2 and Best-of-3 generation, random low-resolution candidate selection, CLIP-based low-resolution ranking with Top-1, Top-2, and Top-4 selection, ImageReward-based Top-1 selection, and combined CLIP–ImageReward ranking. For the LRA methods, eight candidates are generated at 512 × 512 and subsequently used for selection and high-resolution refinement. All final image-quality evaluations are performed at 768 × 768. The results show that all five metric-directed LRA selection strategies achieved higher mean final ImageReward and CLIP scores than the single-pass direct baseline. The largest ImageReward improvement was observed for low-resolution ImageReward Top-1 selection, with a mean difference of 1.0427 over direct generation. The largest CLIP improvement among the LRA methods was observed for CLIP Top-1 selection, with a mean difference of 0.0225. After Holm correction, all five metric-directed LRA methods remained statistically significant against the direct baseline for both evaluation metrics. Random low-resolution selection did not demonstrate statistically significant quality improvements. Direct Best-of-3 also improved both ImageReward and CLIP relative to single-pass direct generation, whereas Direct Best-of-2 did not remain significant after correction. However, direct Best-of-N methods required substantially more generation time. LRA methods required approximately 3.17–3.25 seconds of method-only processing after candidate-pool preparation, but their fully attributed end-to-end time was approximately 37.35–37.42 seconds, compared with 15.98 seconds for direct generation. The shared candidate-pool preparation cost was counted for each LRA method in the per-method end-to-end comparison, although it was physically generated only once per prompt–seed condition. These findings provide exploratory evidence that low-resolution candidate ranking can be associated with improved final image quality after refinement. They do not establish that LRA is universally more computationally efficient or superior to direct generation. The results instead suggest that candidate-selection criteria, candidate-pool preparation costs, refinement behavior, and evaluation-metric dependence are important factors in determining the effectiveness of resolution allocation.