# Flow-Matching Transformer

2 related articles

Subscribe to updates

arXiv Nov 4, 2026
Streaming video-to-audio closes the alignment gap: LiveFoley needs only 0.08 s of frontend look-ahead and runs at 156 ms end-to-end
Real-time video-to-audio has been held back by two bottlenecks: visual frontends shrink their receptive fields to cut look-ahead, and causal diffusion backends denoise each patch in isolation, so both alignment and audio quality trail offline models. LiveFoley uses a trailing sliding window to turn off-the-shelf non-causal segment-level encoders into streaming modules without retraining, cutting frontend look-ahead to 0.08 s, and widens the backend denoising window toward history. On the VGGSound test set, LiveFoley-48k reaches a DeSync of 0.48, on par with offline MMAudio-L at 0.47, with 156 ms end-to-end latency on a single A800 GPU.
Real-time video-to-audio has been held back by two bottlenecks: visual frontends shrink their receptive fields to cut look-ahead, and causal diffusion backends denoise each patch in isolation, so both alignment and audio quality trail offline models. LiveFoley uses a trailing sliding window to turn off-the-shelf non-causal segment-level encoders into streaming modules without retraining, cutting frontend look-ahead to 0.08 s, and widens the backend denoising window toward history. On the VGGSound test set, LiveFoley-48k reaches a DeSync of 0.48, on par with offline MMAudio-L at 0.47, with 156 ms end-to-end latency on a single A800 GPU.
Alibaba Group (China)Nanjing University Monash University Tampere University Tsinghua University
Yusheng Dai（First author）
Video-to-Audio Generation LiveFoley Trailing Sliding Window Autoregressive diffusion Flow-Matching Transformer
arXiv Oct 14, 2026
Pathology generation leaves the 1024-pixel tile behind: coherent gigapixel slides, and synthetic samples lift few-shot classification by up to 18%
Histopathology image generation has stayed at the patch scale, producing tiles of at most 1024×1024 pixels, while clinical text reports and slide-level diagnostic labels act on the whole slide. BigCell trains a flow-matching transformer at 2048×2048 pixels on over 120,000 public whole-slide images and uses an inference-time sliding-window pipeline to generate coherent images up to 32,768×32,768 pixels, outperforming prior tile generators on tile-level and slide-level fidelity metrics. An adaptive scheduling strategy speeds generation by over three times with minimal quality degradation, and synthetic samples improve slide-level few-shot classification by up to 18%.
Histopathology image generation has stayed at the patch scale, producing tiles of at most 1024×1024 pixels, while clinical text reports and slide-level diagnostic labels act on the whole slide. BigCell trains a flow-matching transformer at 2048×2048 pixels on over 120,000 public whole-slide images and uses an inference-time sliding-window pipeline to generate coherent images up to 32,768×32,768 pixels, outperforming prior tile generators on tile-level and slide-level fidelity metrics. An adaptive scheduling strategy speeds generation by over three times with minimal quality degradation, and synthetic samples improve slide-level few-shot classification by up to 18%.
Srikar Yellapragada（First author）
BigCell Flow-Matching Transformer Text-to-Image Synthesis

You’re all caught up

Streaming video-to-audio closes the alignment gap: LiveFoley needs only 0.08 s of frontend look-ahead and runs at 156 ms end-to-end

Pathology generation leaves the 1024-pixel tile behind: coherent gigapixel slides, and synthetic samples lift few-shot classification by up to 18%