Early Time Conditioning for Image Variety

Text-to-image models often produce similar compositions across independent seeds, limiting the range of visual interpretations available for a prompt. We find that the clean-image estimate implied by the first velocity prediction can already reveal a strong layout hypothesis, which subsequent updates tend to refine rather than fundamentally overturn. Merely skipping the earliest updates is sufficient to increase output variety, but this intervention simultaneously changes both the time condition queried by the model and the integration intervals used by the sampler. Controlled decomposition attributes this effect primarily to the model's time query: holding each seed's latent state and integration interval fixed, later-stage time conditioning differentiates first-pass clean-image estimates that otherwise exhibit similar layouts across seeds. Based on this finding, we introduce Early Time Conditioning, which briefly queries the pretrained model with later-stage time conditions while retaining the native integration schedule, text conditioning, and guidance. Across FLUX.1-dev, FLUX.1-schnell, and Qwen-Image on 553 prompts from GenEval with eight outputs each, all four configurations increase spatial and global feature diversity and Vendi Score over native sampling. Dev's single-query rule reaches diversity comparable to Group Inference with near-native HPS, and diversity gains persist among equal-sized sets of GenEval-correct images. The intervention preserves independent sampling and the native model-evaluation count, while requiring neither cross-sample coordination nor internal feature manipulation, making it compatible with standard independent and batched generation pipelines.