Reproduce “SPAE: Spectrally Guided Autoencoder for Pretrained Visual Latents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yibin Huang, Jixiang Hong, Zongzhao Li, Yuhan Dai, Zhibin Wang, Chunwei Wang, Jun Song, Chen Wang and 3 more
SPAE adapts high-dimensional semantic latents from Vision Foundation Models for effective image generation while maintaining their utility for visual understanding. It introduces a spectrally guided autoencoder framework to reorganize the latent space, mitigating spectral mismatch and semantic-detail entanglement, leading to improved reconstruction and competitive generative performance.
Key insights
- A systematic "training-inference spectral mismatch" exists between VFM encoder latents and DiT-generated latents, particularly in high-frequency components, hindering robust generation.
- Semantic and fine-grained detail information are entangled across all channels in high-dimensional VFM latents, indicating a lack of structured organization crucial for effective generative modeling.
- Reorganizing the latent space through bottleneck compression and channel-wise masking can decouple semantic and detail information, creating a more stable and modelable space for diffusion models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.