Lookahead Resolution Allocation (LRA): Second Experimental Report — Evaluating Low-Resolution Candidate Selection, High-Resolution Refinement, and Direct Best-of-N Generation
Year: 2026
Type: preprint
Abstract: High-resolution image generation can be computationally expensive, particularly when multiple candidate images must be generated and evaluated before selecting a final output. This paper presents the second experimental investigation of Lookahead Resolution Allocation (LRA), a candidate-selection and refinement framework that explores whether computational effort can be allocated across different image resolutions. Instead of generating every candidate directly at the target resolution, LRA first generates a pool of low-resolution candidates, evaluates them using image–text similarity and human-preference-oriented scoring models, selects one candidate according to a predefined policy, and subsequently refines that candidate at a higher resolution. The second experiment extends the previous experimental design by introducing two additional direct-generation baselines: Direct Best-of-2 and Direct Best-of-3. These methods generate multiple independent 768 × 768 images and select the candidate with the highest ImageReward score. The experiment therefore compares low-resolution candidate selection and high-resolution refinement against both single-pass direct generation and higher-budget direct Best-of-N generation. A total of 135 experimental conditions were completed using five text prompts, three random seeds, and nine generation or selection methods. Eight 512 × 512 candidates were generated for each prompt–seed combination. Selected low-resolution candidates were refined into 768 × 768 images using an image-to-image diffusion pipeline. All final images were evaluated using CLIP and ImageReward. The descriptive results show that the ImageReward means of the low-resolution ImageReward-based and combined-ranking methods were 0.9776 and 0.9770, respectively, compared with −0.1437 for single-pass direct 768 × 768 generation. However, none of the ImageReward comparisons against the direct baseline remained statistically significant after Holm correction. Four methods—low-resolution CLIP Top-1, CLIP Top-2, ImageReward Top-1, and combined Top-1—showed statistically significant improvements in final CLIP scores after correction. Low-resolution methods required approximately 3.06–3.20 seconds of method-only execution time, compared with 16.18 seconds for direct generation. However, when the shared low-resolution candidate-pool preparation cost was attributed to each method, their end-to-end times were approximately 37.83–37.97 seconds. These findings provide exploratory evidence that low-resolution candidate screening may be a useful allocation strategy under certain evaluation criteria. They do not establish general superiority in perceptual quality or computational efficiency. The experiment is limited by its small prompt set, three random seeds, and the use of ImageReward for both candidate selection and final evaluation. The results are therefore presented as a controlled pilot investigation intended to motivate larger-scale experiments.
Author: Zhang, Jincheng
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225042648