Research map / Image generation / Text-to-image diffusion
Zero-Shot Text-to-Image Generation
Read the paperCheck 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 the large autoregressive text-and-image transformer instead of DF-GAN | MS-COCO captions · MS-COCO captions | human preference for realism: 90.0% | Helped |
| Use the large autoregressive text-and-image transformer instead of DF-GAN | MS-COCO captions · MS-COCO captions | human preference for caption match: 93.3% | Helped |
| Increase the number of candidates used for contrastive reranking (ablation) | DALL·E · MS-COCO | FID: up to 32 | Helped |
| Increase the number of candidates used for contrastive reranking (ablation) | DALL·E · MS-COCO | Inception Score: up to 32 | Helped |
| Apply a slight blur with radius 1 to samples and validation images (ablation) | DALL·E · MS-COCO | FID: about 6 points | Helped |
| Blur samples and validation images with radius greater than or equal to two (ablation) | DALL·E · MS-COCO | Inception Score: greater than or equal to two | Helped |
| Use learned position-specific padding tokens instead of masking padding positions (ablation) | DALL·E transformer · Conceptual Captions out-of-distribution captions | out-of-distribution performance | Helped |
| Use learned position-specific padding tokens instead of masking padding positions (ablation) | DALL·E transformer · Conceptual Captions validation | validation loss | Hurt |
| Increase the dVAE KL weight to β = 6.6 (ablation) | dVAE · image reconstruction | reconstruction error: β = 6.6 | Helped |
| Train the dVAE and transformer jointly instead of using two-stage training (ablation) | DALL·E · ImageNet preliminary experiments | performance | No clear effect |
| Use one shared attention operation for text-to-text, image-to-text, and image-to-image interactions instead of separate operations (ablation) | DALL·E transformer · text-to-image generation | performance | Helped |
| Use a convolutional attention mask in the final self-attention layer (ablation) | DALL·E transformer · text-to-image generation | performance | Helped |
Closest papers by meaning
- Zero-Shot Styled Text Image Generation, but Make It Autoregressive (CVPR 2025 · similarity 0.86)
- Diffusion Self-Distillation for Zero-Shot Customized Image Generation (CVPR 2025 · similarity 0.86)
- Fluid: Scaling Autoregressive Text-to-image Generative Models with Continuous Tokens (ICLR 2025 · similarity 0.86)
- R&B: Region and Boundary Aware Zero-shot Grounded Text-to-image Generation (ICLR 2024 · similarity 0.85)
- TI2V-Zero: Zero-Shot Image Conditioning for Text-to-Video Diffusion Models (CVPR 2024 · similarity 0.85)
- EditAR: Unified Conditional Generation with Autoregressive Models (CVPR 2025 · similarity 0.85)
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.