Research map / Image generation / Autoregressive and latent generation
Taming Sampling Perturbations with Variance Expansion Loss for Latent Diffusion Models
Qifan Li, Xingyu Zhou, Jinhua Zhang, Weiyi You, Shuhang Gu
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 |
|---|---|---|---|
| replace KL with VE loss (ablation) | LDM tokenizer · ImageNet 256 (DiT-B) | FID-10K: 31.93 -> 29.03 | Improved |
| replace KL with VE loss (ablation) | LDM tokenizer · ImageNet 256 (LightningDiT-B) | FID-10K: 22.25 -> 19.70 | Improved |
| replace KL with VE loss (ablation) | LDM tokenizer · ImageNet 256 tokenizer | rFID: 0.55 -> 0.60 | Hurt |
| replace KL with VE loss (ablation) | VA-VAE tokenizer · ImageNet 256 (DiT-B) | FID-10K: 22.27 -> 19.42 | Improved |
| replace KL with VE loss (ablation) | VA-VAE tokenizer · ImageNet 256 (LightningDiT-B) | FID-10K: 19.85 -> 15.50 | Improved |
| replace KL with VE loss (ablation) | VA-VAE tokenizer · ImageNet 256 tokenizer | PSNR: 27.43 -> 26.54 | Hurt |
| train VA-VAE+VE for 16 epochs versus VA-VAE for 50 epochs (ablation) | VA-VAE tokenizer · ImageNet 256 (LightningDiT-B, 160 ep) | FID-10K: 15.82 -> 12.89 | Improved |
| use VE loss instead of raising KL weight to beta=8 (ablation) | VAE tokenizer · ImageNet 256 (no CFG) | FID-10K: 27.54 -> 18.90 | Improved |
| sweep KL weight from 1e-6 to 1 (ablation) | VAE tokenizer · ImageNet 256 (no CFG) | FID-10K: 23.12 -> 23.18 | No clear effect |
| increase KL weight from 1e-6 to 8 (ablation) | VAE tokenizer · ImageNet 256 tokenizer | rFID: 0.39 -> 2.36 | Hurt |
| replace KL with VE loss (ablation) | VAE tokenizer · ImageNet 256 (no CFG) | FID-10K: 23.12 -> 18.90 | Improved |
| fine-tune VA-VAE with VE loss | VA-VAE tokenizer · ImageNet 256 finetuned | rFID: 0.28 -> 0.26 | Improved |
Closest papers by meaning
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.90)
- Smooth Diffusion: Crafting Smooth Latent Spaces in Diffusion Models (CVPR 2024 · similarity 0.90)
- CCS: Controllable and Constrained Sampling with Diffusion Models via Initial Noise Perturbation (NeurIPS 2025 · similarity 0.89)
- What Matters for Diffusion-Friendly Latent Manifold? Prior-Aligned Autoencoders for Latent Diffusion (NeurIPS 2026 · similarity 0.89)
- Consistency Posterior Sampling for Diverse Image Synthesis (CVPR 2025 · similarity 0.89)
- Masked Autoencoders Are Effective Tokenizers for Diffusion Models (ICML 2025 · similarity 0.88)
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.
