High-noise SNR training lifts layout adherence in flow-matching models
Original title: Unlocking Spatial Grounding in Flow-Matching Models via SNR Tuning
Shifting training toward high noise raised synthetic layout adherence from 25.9% to 80.8% and COCO Average Precision from 1.0 to 57.3.
A paper argues that the signal-to-noise ratio schedule determines spatial grounding in flow-matching image models. On a synthetic benchmark of geometric primitives, a centered SNR distribution reached 25.9% layout adherence, while emphasizing high-noise stages reached 80.8%, near a 83.6% VAE reconstruction ceiling. The authors say the shift matters more as resolution and aspect-ratio variance grow, and that conservative schedules such as FLUX.1’s are insufficient. On COCO, distribution tuning alone moved Average Precision from 1.0 to 57.3 without architectural changes.
Unconfirmed: Layout and COCO figures are the authors’ reported results, not an independent evaluation.
- A color-aware layout metric was used to separate structure from texture.
- No auxiliary structural adapters or inference-time steering were required.
- The reported COCO gain comes from SNR distribution tuning alone.
From the original article
1 min read at arxiv.orgDespite 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.
By Martin Thißen · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.