Research map / Label-efficient and robust learning / Self-supervised and contrastive

Generate in Reconstruction Space, Match in Semantic Space: Transport Geometry for One-Step Generation

NeurIPS 2026 · Label-efficient and robust learning › Self-supervised and contrastive · approach: Data-efficient and robust self-supervision · 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.

Shows that one-step generators trained with a Sinkhorn divergence in frozen SSL feature space (on SD-VAE latents) greatly improve ImageNet FID, attributing gains to matching-estimation stability and warning that training on the evaluation featurizer (Inception) enables metric hacking.

Problem

Why SSL features improve generative training is poorly understood; one-step distribution-matching generators have historically lagged strong multi-step models, and the link between the training feature extractor, the evaluation metric (FID), and sample quality is unclear.

Approach

Generate in a VAE latent (generation) space but match generated and real samples with the Sinkhorn divergence in a frozen SSL feature space, with classifier-free guidance built into the training objective; semantic features with lower effective rank give more stable OT barycentric targets, and a diagnostic D̃_N predicts generation quality.

Key results

• Lifting Sinkhorn matching to SSL latent feature spaces reduces FID-Inception from 134.56 (no featurizer) to 3.46 (MAE mask 50%), a 39x reduction.
• Matching stability D̃_128 ranks featurizers consistently with generation quality (MAE 0.164-0.170 vs DINOv3 0.191 vs Inception 0.240).
• Inception-distilled featurizer achieves FID-Inception 4.88 but FD-VAE 510, indicating metric hacking when training features match the evaluation featurizer.

Surprising findings

• The metric-matched featurizer (Inception distillation) is not the best-performing training featurizer for its own metric; MAE wins on FID-Inception.
• Inception-distilled training scores FD-Inc-d 0.17 in its own space while reaching FD-VAE 509, showing near-perfect self-alignment can mask poor generation.

Limitations

• Experiments use a DiT-B/2 generator; scaling to larger architectures is untested.
• Reported FID gap to strongest one-step and multi-step models remains; competitive but not state-of-the-art.
• Geometric alignment R^2 is a proxy computed on a small sample (2048 latents) and only post-hoc diagnostic for generator-level misalignment.

Experimental setup

Datasets: ImageNet

Benchmarks: ImageNet 256x256 class-conditional generation

Metrics: FID-Inception (FD) FD-DINOv3 FD-VAE effective rank D̃_N matching stability

Baselines: No featurizer (VAE latent matching) Inception distillation DINOv3 distillation drifting

Builds on: Sinkhorn divergence REPA drifting generative models DiT Latent diffusion

Closest papers by meaning

• Can Generative Models Improve Self-Supervised Representation Learning? (AAAI 2025 · similarity 0.89)
• Boosting Generative Image Modeling via Joint Image-Feature Synthesis (NeurIPS 2025 · similarity 0.86)
• Look Ahead or Look Around? A Theoretical Comparison Between Autoregressive and Masked Pretraining (ICML 2024 · similarity 0.85)
• Blockwise Flow Matching: Improving Flow Matching Models For Efficient High-Quality Generation (NeurIPS 2025 · similarity 0.85)
• Beckmann Transport Models: From Autonomous Flows to One-Step Maps (NeurIPS 2026 · similarity 0.85)
• When Worse is Better: Navigating the Compression Generation Trade-off In Visual Tokenization (NeurIPS 2025 · similarity 0.85)