Research map / Image generation / Autoregressive and latent generation
Taming Transformers for High-Resolution Image Synthesis
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 |
|---|---|---|---|
| Use context-rich VQGAN codes instead of direct pixel modeling | Transformer image synthesis on CIFAR10 · CIFAR10 | FID: 18.63% | Helped |
| Use context-rich VQGAN codes instead of direct pixel modeling | Transformer image synthesis on CIFAR10 · CIFAR10 | sampling speed: 14.08× faster sampling | Helped |
| Use row-major ordering rather than the tested alternative orderings (ablation) | VQGAN + transformer · ImageNet validation split | negative log-likelihood: 4.767 vs 4.889, 4.810, 5.015, 4.812, 4.901 | Helped |
| Use subsample ordering instead of row-major ordering (ablation) | VQGAN + transformer · ImageNet validation split | negative log-likelihood: 4.767 -> 5.015 | Hurt |
| Increase the first-stage context through larger downsampling factor (ablation) | VQGAN + transformer · FacesHQ | sample fidelity | Helped |
| Use a small context factor rather than f=16 (ablation) | VQGAN + transformer · FacesHQ | sample coherence | Hurt |
| Use f=8 rather than f=16 (ablation) | VQGAN + transformer · FacesHQ | sample fidelity | Hurt |
| Select checkpoints by best training NLL rather than best validation NLL (ablation) | VQGAN + transformer · CelebA-HQ | FID: 10.2 -> 3.86 | Helped |
| Select checkpoints by best training NLL rather than best validation NLL (ablation) | VQGAN + transformer · FFHQ | FID: 9.6 -> 2.68 | Helped |
| Select checkpoints by best training NLL rather than best validation NLL (ablation) | VQGAN + transformer · CelebA-HQ and FFHQ | overfitting | Hurt |
| Use a larger compression factor for VQGAN codes (ablation) | VQGAN + transformer · Image synthesis | transformer sequence modeling efficiency | Helped |
| Increase compression beyond the dataset-specific reconstruction limit (ablation) | VQGAN · Image reconstruction | reconstruction fidelity | Hurt |
Closest papers by meaning
- Win-Win: Training High-Resolution Vision Transformers from Two Windows (ICLR 2024 · similarity 0.90)
- An Image is Worth More Than 16x16 Patches: Exploring Transformers on Individual Pixels (ICLR 2025 · similarity 0.88)
- Scalable High-Resolution Pixel-Space Image Synthesis with Hourglass Diffusion Transformers (ICML 2024 · similarity 0.88)
- LaVin-DiT: Large Vision Diffusion Transformer (CVPR 2025 · similarity 0.87)
- Viewing Transformers Through the Lens of Long Convolutions Layers (ICML 2024 · similarity 0.87)
- UltraPixel: Advancing Ultra High-Resolution Image Synthesis to New Peaks (NeurIPS 2024 · similarity 0.86)
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.