研究显示调整 SNR 分布可显著提升 flow-matching 模型空间布局能力
Unlocking Spatial Grounding in Flow-Matching Models via SNR Tuning
一群研究者发现不用改架构，只要把 flow-matching 模型的训练噪声分布往高噪阶段偏移，布局遵循度就能从 25.9% 涨到 80.8%，训练文生图模型的值得看看。
一篇 arXiv 论文发现 flow-matching 模型中文本编码器本已具备表示空间坐标的潜在能力，问题出在信噪比（SNR）训练分布上。在使用几何图形合成基准测试时，以居中 SNR 分布训练的模型布局遵循度只有 25.9%，而将训练重心移向高噪声阶段后提升到 80.8%，接近 VAE 重建上限 83.6%。在 COCO 真实图像上，Average Precision 从 1.0 提升到 57.3。论文还指出 FLUX.1 采用的保守噪声调度不足以实现精确布局，而这一改进无需修改模型架构。
Unlocking Spatial Grounding in Flow-Matching Models via SNR Tuning
Despite the success of foundation models in image synthesis, achieving fine-grained spatial control remains a significant challenge, typically requiring auxiliary structural adapters or complex inference-time steering. While modern vision-language (VLM) text encoders possess the latent capacity to represent spatial coordinates, generative models often fail to leverage this information. We demonstrate that the signal-to-noise ratio (SNR) distribution is a critical determinant of spatial grounding in flow-matching models. Using a novel synthetic benchmark of geometric primitives, we isolate structural mechanics from textural complexity inherent in natural imagery and introduce a rigorous color-aware layout metric to quantify spatial fidelity. We show that training on a centered SNR distribution results in poor layout adherence (25.9%). Instead, we identify high-noise regimes as a "critical learning phase" for spatial grounding. Shifting the training focus toward these stages increases layout adherence to 80.8%, effectively reaching the 83.6% reconstruction ceiling of the underlying variational autoencoder. We find that the necessity for this shift scales with resolution and aspect ratio variance, and that the conservative noise schedules employed by models like FLUX.1 are insufficient for precise layout adherence. Finally, we show that these findings translate to real-world imagery via the COCO dataset (1.0 vs. 57.3 Average Precision), where state-of-the-art layout adherence is unlocked through distribution tuning alone, without architectural modifications.