Research map / Probabilistic and scientific modeling / Variational and generative sampling

Rethinking Vector Field Learning for Generative Segmentation

NeurIPS 2026 · Probabilistic and scientific modeling › Variational and generative sampling · approach: Flow matching and flow sampling · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Reshapes the flow matching velocity target with a detached, distance-aware potential-gradient correction (attraction plus repulsion) on top of a pixel neural field (PixNerd), with Kronecker-style quasi-random class-centroid encoding, for generative semantic segmentation.

Problem

Flow-matching-based diffusion segmentation converges slowly and separates classes poorly in high-cardinality settings, due to gradient vanishing near class centroids and lack of repulsion from competing classes.

Approach

Augment the ground-truth velocity with a stop-gradient correction derived from the gradient of a softmax-over-negative-log-distance potential to centroids, giving attractive and repulsive forces while keeping the standard diffusion training framework; pair it with quasi-random centroid encoding and end-to-end pixel neural field decoding.

Key results

• FlowSeg reaches 47.1 mIoU on ADE20K, above the best discriminative baseline MaskFormer (46.7) and the diffusion baselines InstructDiffusion (33.6) and PixWizard (32.8).
• FlowSeg reaches 44.9 mIoU on COCO-Stuff, above SegFormer (44.6) and SymmFlow (39.6).
• In ablations on ADE20K, REPA raises mIoU from 29.8 to 32.7, and 10 Euler sampling steps give the peak of 47.1 mIoU.

Surprising findings

• Annealing the correction term over time degrades performance (24.65 vs 28.15 mIoU for static v-pred), which the authors attribute to weakened discriminative potential.
• Performance peaks at 10 sampling steps and declines with more steps (45.78 at 100 steps).

Limitations

• Performance peaks at 10 sampling steps and declines with more steps, indicating sensitivity to the sampler configuration.
• Results are on only two semantic segmentation datasets, both high-cardinality.
• Relies on a hyperparameter temperature and a choice of distance transform T; gentler alternatives were found to underperform.

Experimental setup

Datasets: ADE20K COCO-Stuff ImageNet-1k

Benchmarks: ADE20K COCO-Stuff

Metrics: mIoU

Baselines: DeeplabV3+ OCRNet SegFormer MaskFormer InstructDiffusion PixWizard SymmFlow

Builds on: Rectified Flow PixNerd REPA flow matching

Closest papers by meaning

• Diff2Flow: Training Flow Matching Models via Diffusion Model Alignment (CVPR 2025 · similarity 0.89)
• Value Gradient Guidance for Flow Matching Alignment (NeurIPS 2025 · similarity 0.87)
• Differential Vector Erasure: Unified Training-Free Concept Erasure for Flow Matching Models (NeurIPS 2026 · similarity 0.87)
• Seg4Diff: Unveiling Open-Vocabulary Semantic Segmentation in Text-to-Image Diffusion Transformers (NeurIPS 2025 · similarity 0.87)
• EmerDiff: Emerging Pixel-level Semantic Knowledge in Diffusion Models (ICLR 2024 · similarity 0.87)
• Unleashing the Potential of the Diffusion Model in Few-shot Semantic Segmentation (NeurIPS 2024 · similarity 0.86)