Research map / Image generation / Autoregressive and latent generation
Unifying CNNs and ViTs for Learning-Efficient and Scalable Variational AutoEncoder
Zhiying Lu, Shang Chai, Chuanbin Liu, Litong Gong, Pandeng Li, Tiezheng Ge, Hongtao Xie
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use RoPE only for positional encoding instead of absolute position embedding (ablation) | ViT-VAE · Extrapolation 256 to 512 | reconstruction quality | Improved |
| add a positional encoding (RoPE) to a ViT-VAE without any (ablation) | ViT-VAE · Early training on ImageNet-1k | validation reconstruction loss | Improved |
| replace patch embedding and pixel-shuffle with sequential overlapping convolutions (SeqConv) (ablation) | ViT w/ RoPE · Early training on ImageNet-1k | validation reconstruction loss | Improved |
| add a multi-stage design with ResBlock front-end (ablation) | ViT w/ RoPE · Early training on ImageNet-1k | validation reconstruction loss | Improved |
| add depth-wise convolution FFN (ablation) | ViT w/ RoPE w/ Multi-Stage · Early training on ImageNet-1k | validation reconstruction loss | Improved |
| add full convolution FFN (ablation) | ViT w/ RoPE w/ Multi-Stage · Early training on ImageNet-1k | validation reconstruction loss | Improved |
| replace FullConv-FFN with MoE-FFN (ablation) | TransVAE · Early training on ImageNet-1k | validation reconstruction loss | Hurt |
| add RMSNorm before QKV projections (ablation) | TransVAE · long training at scale | training stability | Improved |
| scale TransVAE from Tiny to Large | TransVAE · ImageNet-1k val 256x256 (f16d32) | PSNR: 29.08 -> 29.78 | Improved |
| scale TransVAE from Tiny to Large | TransVAE · ImageNet-1k val 1024x1024 (f16d32) | PSNR: 36.38 -> 38.25 | Improved |
| add VF alignment loss | TransVAE-T · ImageNet-1k val 256x256 (f16d32) | PSNR: 29.08 -> 28.76 | Hurt |
| add VF alignment loss | TransVAE-T · ImageNet-1k val 1024x1024 (f16d32) | rFID: 0.08 -> 0.24 | Hurt |
Closest papers by meaning
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.91)
- SRA 2: Variational Autoencoder Self-Representation Alignment for Efficient Diffusion Training (CVPR 2026 · similarity 0.90)
- NVAE: A Deep Hierarchical Variational Autoencoder (NeurIPS 2020 · similarity 0.89)
- CV-VAE: A Compatible Video VAE for Latent Generative Video Models (NeurIPS 2024 · similarity 0.89)
- LaVin-DiT: Large Vision Diffusion Transformer (CVPR 2025 · similarity 0.89)
- Improved Video VAE for Latent Video Diffusion Model (CVPR 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
