Research map / Image generation / Autoregressive and latent generation
High-Resolution Image Synthesis With Latent Diffusion Models
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Train diffusion in a perceptually compressed latent space rather than pixel space | LDM · ImageNet, class-conditional synthesis after 2M training steps | FID: 38 | Helped |
| Use a moderate downsampling factor (LDM-4 to LDM-16) instead of pixel-level or very strong compression (ablation) | LDM · ImageNet, class-conditional synthesis | FID | Helped |
| Use a small downsampling factor (LDM-1 or LDM-2) (ablation) | LDM · ImageNet, class-conditional synthesis | sample quality | Hurt |
| Use an overly large downsampling factor (ablation) | LDM · ImageNet, class-conditional synthesis | sample fidelity | Hurt |
| Use classifier-free guidance (ablation) | LDM-KL-8 text-to-image model · MS-COCO validation | sample quality | Helped |
| Use latent diffusion rather than pixel-based diffusion | LDM inpainting · Places validation split | FID: 1.6× | Helped |
| Use latent diffusion rather than pixel-based diffusion | LDM inpainting · Places validation split | sampling throughput: 2.7× | Helped |
| Add attention to the first-stage model (ablation) | LDM-4 inpainting · Places inpainting | FID | Helped |
| Fine-tune the larger VQ-regularized model without first-stage attention at 512² (ablation) | LDM inpainting · Places, 512² | FID | Helped |
| Replace bicubic-only conditioning degradation with diverse BSR degradation (ablation) | LDM super-resolution · Real-world images | image sharpness | Helped |
| Use bicubic-only degradation rather than diverse degradation (ablation) | LDM super-resolution · Images outside the training preprocessing pipeline | generalization | Hurt |
| Rescale latent components by their standard deviation (ablation) | LDM convolutional sampling · Semantic image synthesis at resolutions above 256² | sample quality | Helped |
Closest papers by meaning
- DiffuseHigh: Training-Free Progressive High-Resolution Image Synthesis Through Structure Guidance (AAAI 2025 · similarity 0.92)
- Boosting Generative Image Modeling via Joint Image-Feature Synthesis (NeurIPS 2025 · similarity 0.92)
- ScaleDiff: Higher-Resolution Image Synthesis via Efficient and Model-Agnostic Diffusion (NeurIPS 2025 · similarity 0.92)
- Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (CVPR 2025 · similarity 0.92)
- Simpler Diffusion: 1.5 FID on ImageNet512 with Pixel-space Diffusion (CVPR 2025 · similarity 0.91)
- Masked Autoencoders Are Effective Tokenizers for Diffusion Models (ICML 2025 · similarity 0.91)
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.