Composite flow matching improves image representation and generation
Representation Learning via Composite Subspace Flow Matching
Computer Vision and Pattern RecognitionArtificial Intelligence
Summary
Understanding images so computers can recognize and recreate them is tricky because methods usually focus on either the important content or the exact details, but not both. The authors introduce a new approach called composite subspace flow matching that teaches a system to separate lasting content from temporary style changes like lighting or color. This way, the system learns useful features for recognizing images while also being able to generate realistic pictures. Tests on popular image sets show the method works well with different types of image models.
What this means in practice
- •For computer vision engineers: Train models that both classify images accurately and generate photo-realistic variations by separating content from style changes.
- •For digital content creators: Create smooth image transformations and interpolations to produce realistic edits and animations from learned representations.
Authors
José Afonso, Mário Figueiredo
Abstract
Distilling semantic meaning from raw image data remains a challenge in machine learning. Current self-supervised paradigms face a trade-off: joint-embedding architectures produce relevant features by aggressively discarding intra-class variance, irreversibly spatial and reconstruction. Conversely, generative models preserve spatial details but struggle to isolate semantics, as their optimization is dominated by pixel variations. To bridge this divide, we propose composite subspace flow matching (CSFM). Rather than corrupting data with unstructured isotropic noise, CSFM formulates representation learning as the variance-minimizing inversion of a deterministic degradation curriculum across spectral, color, and illumination subspaces. By bounding trajectory variance with hard temporal style truncation, a capacity-constrained encoder is driven to ignore transient style artifacts and encode persistent content. Empirical evaluations across CIFAR-10, CIFAR-100, and STL-10 show that CSFM yields competitive discriminative transferability while preserving generative synthesis across both convolutional and vision transformer architectures. By avoiding the collapse pathologies of non-generative methods, CSFM forms a smooth latent space that outperforms standard optimal transport flows on linear probing and few-shot classification, while enabling coherent synthesis and trajectory interpolation. These results confirm that steering the physical dynamics of generative flows offers a scalable paradigm for learning representations that are simultaneously linearly separable and generatively invertible.