CoVA: Composing VAE Representations for Generative Learning
Representation alignment with pretrained visual encoders improves diffusion transformer training but requires auxiliary encoders and extra training overhead.
ProofPaper ↗
Key points
- Recent approaches instead exploit representations from the generative model's own variational autoencoder (VAE).
- We introduce CoVA (Composed VAE Representations), a lightweight framework that composes VAE features online into stage-wise guidance for flow-matching training.
- A small latent composer is jointly trained with class supervision, masked latent reconstruction, and horizontal-flip consistency, yielding semantically discriminative, spatially structured dense features and sample-dependent class posteriors.
- CoVA aligns intermediate SiT features with the composed representations early in training, then removes the increasingly detrimental alignment and reuses detached composer posteriors for sample-dependent soft label conditioning.
Sources (1)
- [1]CoVA: Composing VAE Representations for Generative LearningarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 9, 10:03 AM
Representation alignment with pretrained visual encoders improves diffusion transformer training but requires auxiliary encoders and extra training overhead.
Recent approaches instead exploit representations from the generative model's own variational autoencoder (VAE).
Extractive summary: sentences quoted from the sources.
Before this
- Oct 9, 2026Destination-Conditioned Adaptation with Semantic-Geometric Alignment for Generative Inbetweening
- Oct 9, 2026Qwen/Qwen-Image-2.1-Turbo
- Oct 9, 2026Learning Representations to Guide Pixel Diffusion Transformers
- Oct 8, 2026Few-Step Generation via Data-Space Iteration
- Oct 8, 2026Normalizing Trajectory Models
- Oct 7, 2026Velocity Scaling in Flow Matching
