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

Stochastic Interpolants via Conditional Dependent Coupling

NeurIPS 2026 · Probabilistic and scientific modeling › Variational and generative sampling · approach: Regularizing VAE latent information

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 CDC-FM, a multi-stage pixel-space flow matching framework where each resolution stage is a stochastic interpolant conditioned on an upsampled lower-resolution image plus small noise, all parameterized by one DiT trained end-to-end.

Problem

Pixel-space image generation is computationally expensive, while latent (VAE) models lose information and cannot be trained end-to-end; existing multi-stage cascades use disentangled stage-specific models that hurt distribution learning and end-to-end optimization.

Approach

Conditional dependent coupling: the source for each refinement stage is the upsampled-and-downsampled ground-truth image plus stage-decreasing Gaussian noise, so each stage's interpolant transports a structured prior to the target; a single shared DiT with a resolution embedding models all stages.

Key results

• On ImageNet-1k 256x256, CDC-FM with 6x4x2 NFE reaches FID 1.81 and IS 304.3 (677M params).
• On ImageNet-1k 512x512, CDC-FM (684M) reports FID 2.65 and IS 307.4, compared with DiT-XL/2 at FID 3.04.
• Inference time reported to scale linearly with image size across 64 to 512 resolutions.

Limitations

• Theoretical cost comparison relies on strong assumptions (orthogonal detail energies, NFE proportional to transport cost).
• Source code is promised only upon acceptance; no release at submission.
• Stage noise schedule and diminish factor gamma require tuning; FID worsens beyond a tipping point.

Experimental setup

Datasets: ImageNet-1k

Benchmarks: ImageNet-1k 256x256 ImageNet-1k 512x512

Metrics: FID IS precision recall inference speed

Baselines: DiT-XL/2 SiT-XL/2 PixelFlow VAR LlamaGen MAR FractalMAR PaGoDA VDM++ SiD2

Builds on: stochastic interpolant flow matching DiT classifier-free guidance PixelFlow

Closest papers by meaning

• Stochastic Interpolants with Data-Dependent Couplings (ICML 2024 · similarity 0.89)
• A Simple Approach to Unifying Diffusion-based Conditional Generation (ICLR 2025 · similarity 0.88)
• Multitask Learning with Stochastic Interpolants (NeurIPS 2025 · similarity 0.88)
• Boosting Generative Image Modeling via Joint Image-Feature Synthesis (NeurIPS 2025 · similarity 0.88)
• Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.88)
• Multimarginal Generative Modeling with Stochastic Interpolants (ICLR 2024 · similarity 0.88)