Lookahead Noise Intervention Optimization (LNIO): Second Experimental Report — A Controlled Stochastic Exploration of Preview-Ranked Sampling Policies in Diffusion Models

Diffusion models generate images through iterative denoising, and the stochasticity introduced during sampling can influence the resulting image. Lookahead Noise Intervention Optimization (LNIO) investigates whether evaluating candidate stochastic actions through a short preview of their downstream generation trajectories can provide useful information for selecting sampling interventions. Following an initial controlled pilot study, this second experimental report examines a broader set of stochastic selection policies, with particular attention to preview-ranked candidate selection and the effect of suppressing the selection probability assigned to the second-ranked candidate. The experiment evaluates 14 methods across 22 prompts and five random seeds, producing 1,540 completed generation conditions. The methods include a deterministic baseline, a heuristic adaptive baseline, three Random Direct methods, and nine preview-based Random Lookahead configurations. Random Direct selects among four candidate DDIM stochasticity parameters uniformly without preview evaluation. Random Lookahead evaluates all four candidates, ranks them according to preview CLIP similarity, and samples from a rank-dependent probability distribution. Three probability configurations are examined at decision intervals of 8, 4, and 2 denoising steps. ImageReward is treated as the primary evaluation metric, while CLIP similarity is used as a secondary metric. The Adaptive baseline obtains the highest mean ImageReward score (-0.4004), compared with -0.5412 for the deterministic baseline. Among the Random Lookahead configurations, E4 Bias10 obtains the highest mean ImageReward score (-0.4534), corresponding to a mean difference of +0.0878 relative to the deterministic baseline. However, none of the reported primary paired comparisons remains statistically significant after Holm correction. The E4 Bias10 configuration also requires approximately 1.96 times the baseline runtime and 14 rather than 8 UNet forward calls per generation. The findings provide a broader empirical examination of preview-based stochastic selection, but do not establish a statistically significant advantage over the evaluated baselines. In particular, the results suggest that the relationship between preview ranking, probability allocation, decision frequency, evaluation metrics, and computational overhead requires further investigation. This report therefore presents LNIO as an experimental framework under evaluation rather than as a method with demonstrated superiority.