Research map / Image generation / Autoregressive and latent generation
DA-VAE: Plug-in Latent Compression for Diffusion via Detail Alignment
Xin Cai, Zhiyuan You, Zhoutong Zhang, Tianfan Xue
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 |
|---|---|---|---|
| add detail alignment loss to detail channels (ablation) | DA-VAE · ImageNet 512 ablation | FID-10k: 16.37 -> 9.27 | Improved |
| zero-initialize new patch embedder and output layer (ablation) | DA-VAE fine-tuning · ImageNet 512 ablation | FID-10k: 29.73 -> 9.27 | Improved |
| add cosine-annealed loss weight schedule on detail channels (ablation) | DA-VAE fine-tuning · ImageNet 512 ablation | FID-10k: 9.80 -> 9.27 | Improved |
| increase alignment weight from 0.5 to 1.0 (ablation) | DA-VAE · ImageNet 512 ablation | rFID: 0.47 -> 0.63 | Hurt |
| increase alignment weight from 0.5 to 1.0 (ablation) | DA-VAE · ImageNet 512 ablation | FID-10k: 9.27 -> 9.23 | No clear effect |
| increase alignment weight from 0 to 0.1 (ablation) | DA-VAE · ImageNet 512 ablation | FID-10k: 16.37 -> 9.58 | Improved |
| use DA-VAE latent (f32c128) instead of VA-VAE (f16c32) | DiT-XL from scratch · ImageNet 512 | FID-10k: 44.65 -> 31.51 | Improved |
| fine-tune with DA-VAE instead of DC-Gen fine-tuning | LightningDiT-XL · ImageNet 512 (80 epochs) | FID-50k w/ CFG: 2.22 -> 1.68 | Improved |
| fine-tune SD3.5M with DA-VAE at 32x32 tokens | SD3.5-medium · MJHQ-30K 1024x1024 | throughput: 0.25 -> 1.03 img/s | Improved |
| fine-tune SD3.5M with DA-VAE at 32x32 tokens | SD3.5-medium · MJHQ-30K 1024x1024 | FID: 10.31 -> 10.91 | Mixed |
Closest papers by meaning
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.92)
- DC-SAE: Deep Compression Semantic Autoencoder for Faster Diffusion Convergence (NeurIPS 2026 · similarity 0.91)
- OSCAR: One-Step Diffusion Codec Across Multiple Bit-rates (NeurIPS 2025 · similarity 0.90)
- REPA Works Until It Doesn’t: Early-Stopped, Holistic Alignment Supercharges Diffusion Training (NeurIPS 2025 · similarity 0.90)
- Representation Entanglement for Generation: Training Diffusion Transformers Is Much Easier Than You Think (NeurIPS 2025 · similarity 0.89)
- Masked Autoencoders Are Effective Tokenizers for Diffusion Models (ICML 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.
