SNR tuning improves spatial control in image synthesis models
Unlocking Spatial Grounding in Flow-Matching Models via SNR Tuning
Computer Vision and Pattern RecognitionMachine Learning
Summary
Getting computer models to draw images with accurate layouts is hard. The authors show that when training these image models, focusing on certain noisy stages helps the model better understand and recreate the spatial arrangement of objects. They tested this on simple shapes and real photos, showing much better alignment between the intended layout and the generated image. This approach improves control without needing extra modifications to the model design.
What this means in practice
- •For image generation engineers: Improve the spatial accuracy of generative image models by adjusting training noise schedules without added architectural components.
- •For computer vision product teams: Create tools producing images with precise object layouts from text without complex run-time control methods.$Commercial implications: Enables commercial image synthesis software to offer better spatial layout fidelity, enhancing product quality and user control.
Authors
Martin Thißen, Elke Hergenröther
Abstract
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.