Research map / Image generation / Autoregressive and latent generation

STREAM: Stochastic Riemannian Flow Matching with Anisotropic Decoder for Digital Histopathology Image Generation

NeurIPS 2026 · Image generation › Autoregressive and latent generation · approach: High-resolution diffusion synthesis · 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.

Proposes STREAM, a Riemannian flow matching framework that generates histopathology images in frozen UNI patch-token latent space on the hypersphere, using a stochastic bridge perturbation for the DiT and a spectrally-shaped anisotropic decoder noise.

Problem

Unconditional histopathology image generation: existing state-of-the-art models (ZoomLDM, PixCell) rely on VAE latents conditioned on VFM embeddings, which yields conditioning-dominated diversity and cannot sample without a conditioning signal.

Approach

Uses histopathology VFM (UNI) patch tokens directly as the generative latent on the product hypersphere; trains a DiT with a bridge-type stochastic perturbation of SLERP paths and a chord loss, and trains a decoder with noise covariance shaped by the left singular basis of the DiT velocity-field Jacobian (high noise on low-response directions).

Key results

• STREAM achieves gFID 6.16 on BRCA and 7.68 on COADREAD, best among all compared methods including ZoomLDM (7.43 / 8.09) and RAE (15.84 / 19.64).
• Ablation on BRCA: bridge plus anisotropic decoder reduces rFID to 3.52 and gFID to 6.86, versus 6.51/9.07 for neither component.
• Linear probe AUROC on SPIDER-breast is 0.995 for UNI versus 0.640 for PixCell VAE and 0.624 for ZoomLDM VAE.

Surprising findings

• Bridge alone with isotropic decoder slightly degrades gFID (9.07 to 9.37), but combined with anisotropic decoder the two components act superadditively.
• STREAM is not state of the art on reconstruction: SVG attains best rFID on both datasets.

Limitations

• Theoretical rectifiability guarantee holds only at the population level and relies on Riemannian extension treated as an assumption.
• Mechanism behind the superadditive effect is mostly empirical.
• Reconstruction quality trails SVG and, on COADREAD, PixCell and ZoomLDM.

Experimental setup

Datasets: TCGA-BRCA TCGA-COADREAD SPIDER-breast SPIDER-skin

Benchmarks: TCGA-BRCA TCGA-COADREAD SPIDER-skin

Metrics: FID KID CMMD FvD KvD vMMD LPIPS gFID rFID

Baselines: ZoomLDM PixCell RAE SVG REPA-E

Builds on: flow matching rectified flow RAE LightningDiT VQGAN

Closest papers by meaning

• Dynamic Entity-Masked Graph Diffusion Model for Histopathology Image Representation Learning (AAAI 2025 · similarity 0.88)
• Semantic and Visual Crop-Guided Diffusion Models for Heterogeneous Tissue Synthesis in Histopathology (NeurIPS 2025 · similarity 0.88)
• ZoomLDM: Latent Diffusion Model for Multi-scale Image Generation (CVPR 2025 · similarity 0.87)
• MedFlowSeg: Flow Matching for Medical Image Segmentation with Frequency-Aware Attention (NeurIPS 2026 · similarity 0.86)
• STimage-1K4M: A histopathology image-gene expression dataset for spatial transcriptomics (NeurIPS 2024 · similarity 0.86)
• Learned Representation-Guided Diffusion Models for Large-Image Generation (CVPR 2024 · similarity 0.86)