Lookahead Resolution Allocation for Text-to-Image Diffusion Models: A Pilot Study of 512×512 Candidate Screening and 768×768 Refinement

Generating high-resolution images with text-to-image diffusion models can require substantial computational resources. An alternative is to generate multiple candidates at an intermediate resolution, evaluate them using image–text alignment or learned image-quality signals, and refine a selected candidate to a higher resolution. This paper investigates this strategy through a pilot study of Lookahead Resolution Allocation (LRA), a candidate-screening and refinement framework that allocates image-generation computation across two resolutions. The experiment compares direct image generation at 768×768 resolution with six screening strategies. For each prompt–seed combination, eight candidate images are generated at 512×512 resolution. A candidate is then selected using random selection, CLIP-based Top-1, Top-2, or Top-4 selection, ImageReward-based Top-1 selection, or a combined rank-based strategy. The selected candidate is refined to 768×768 using an image-to-image diffusion pipeline. All final images are evaluated using CLIP and ImageReward. The experiment includes five text prompts, three random seeds, and seven methods, producing 105 completed experimental conditions. Across the six screening strategies, the mean final ImageReward and CLIP scores were higher than those obtained through direct 768×768 generation under the tested configuration. ImageReward-based Top-1 selection achieved the highest mean final ImageReward score (0.9776), while CLIP-based Top-1 selection achieved the highest mean final CLIP score (0.3624). However, none of the ImageReward comparisons reached statistical significance after Holm correction when using prompt-level clustering. Four screening strategies showed statistically significant improvements in final CLIP scores after correction. These results require cautious interpretation because of the small number of prompt clusters and the reuse of evaluation models during candidate selection. Although the measured method-only runtime of the screening-and-refinement strategies was approximately 3.05–3.17 seconds, compared with 15.75 seconds for direct generation, this measurement excludes the shared candidate-generation and scoring costs. The findings provide preliminary evidence that intermediate-resolution candidate evaluation may be useful for guiding subsequent high-resolution refinement. Nevertheless, the experiment does not establish general candidate-ranking validity, end-to-end computational efficiency, or superiority over direct high-resolution generation. Further experiments involving full candidate-level refinement, independent evaluation metrics, larger prompt sets, and matched computational budgets are required.