Latent Barrier Steering adds hierarchical safety to generative planners
On navigation and robot end-effector tasks, LBS kept safety while improving goal reaching versus path-level filters.
Stories per day · 30 days
7 RSS
Relatedrobotics5generative-models4computer-vision2diffusion2dinov22forecasting2foundation-model2image-restoration2pathfm2streaming2time-series2vla23d-generation1action-chunk1action-learning1adversarial-distillation1adversarial-robustness1anomaly-detection1
71 stories, newest first · one card per story
Stories tagged #flow-matching
LipsFlower constrains flow matching to stabilize generative image restoration
LipHDiT uses spectral normalization, bounded modulation and finite-radius gain penalties, tested on blind face restoration.
Patch-wise PCA links v-prediction failure to a bandwidth bottleneck
A Raw-injected Transformer restores effective v-prediction and is competitive on ImageNet 256×256.
JoVox jointly models voxel hierarchy with discrete flow matching
Authors report 8x training throughput, 1/15 forward FLOPs, and 11% lower Chamfer distance than TRELLIS.2.
FlowMoP generates stochastic multi-person motion prediction
FlowMoP uses flow matching for diverse multi-agent motions achieving SOTA on CMU-Mocap and 3DPW.
FlowAlign aligns source representations to target space using flow matching
FlowAlign repurposes conditional flow matching to preserve paired correspondences while matching global target distribution structure.
FewStepRL adapts diffusion RL objectives to few-step samplers
The method recasts FlowGRPO, AWM, DiffusionNFT, DiffusionOPSD and ReFL so rollouts no longer need a many-step sampler.
Ruby-OPD enhances pronunciation controllability for flow-matching TTS
Ruby-OPD framework improves pronunciation accuracy on raw orthographic text for Japanese flow-matching TTS.
VA-TDM boosts few-step diffusion models
VA-TDM uses velocity adversarial distillation to reach state-of-the-art four-step text-to-image and text-to-video generation.
SphereFlow imputes missing modalities via hyperspherical transport
SphereFlow projects modalities to shared hypersphere for short-range cross-modal flow, achieving state-of-the-art on BraTS MRI imputation and closing 80% of segmentation gap.
RynnVLA-Latent pretrains a 4B robot policy on 3D latent actions
A 4B policy pretrained on 745.52M descriptors from 30.1K hours of video posts strong LIBERO, VLABench, and RoboTwin scores.
Flownomaly recovers normal features for few-shot anomaly detection
In the 1-shot setting, image-level AUROC reached 97.8% on MVTec-AD and 95.3% on VisA.
GLOW-FM enhances low-light images with one latent flow step
The small variant uses a frozen Tiny-VAE and Terminal Velocity Matching at 34 GFLOPs and one function evaluation.
Velocity Ambiguity Profiles split flow-matching loss by interpolation time
Guided time sampling cuts DINOv2 feature velocity loss 6.78% and Fréchet distance 23.6%, with no class-conditional KID gain.
High-noise SNR training lifts layout adherence in flow-matching models
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.
StyleStream 2.0 cuts streaming voice conversion latency to 520 ms
A text-conditioned style embedding supports prompt-based voice design and instruction-based editing of one attribute.
DeOverlay reconstructs clean video by anchoring flow to source latents
DVTE-Bench covers 1,631 paired videos; DeOverlay lowers Mask LPIPS 58.3% and Mask VFID 70.3% versus the strongest baseline.
StreamingSwap aims at real-time identity-preserving video face swapping
Authors report competitive identity transfer at 20 FPS on one RTX 4090, stable on videos up to 10 minutes.
VLA-Rolling reuses action tails to cut VLA denoising steps
Across 16 frozen model-benchmark pairs it lowered chunk inference cost, with up to 1.48 times speedup and higher success in 14.
MambaAct swaps pi0.5 transformer action expert for compact Mamba
The public model keeps the pi0.5 vision-language encoder, shrinks the action expert 10 times, and raises mean LIBERO success by 1.08 points.
PathFM: Fast Generative Forecasting of Multivariate Trajectories
PathFM generates full forecast trajectories with fast flow-matching sampling and leads a new path-score benchmark.
PathFM: Fast Generative Forecasting of Multivariate Trajectories
PathFM uses flow matching to forecast full multivariate trajectories and beats baselines on a new path-score benchmark.
Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy — Jason Ma, Dyna Robotics
Dyna Robotics details Dyna-2, a world-action model trained on a million hours and deployed on working robots.
5
Self-Conditioning Enables Deep Sequential Reasoning in Flow Matching
Researchers show self-conditioning lets flow-matching models gain deeper visual reasoning from additional sampling steps.