Research map / Image generation / Text-to-image diffusion
StyleGAN-T: Unlocking the Power of GANs for Fast Large-Scale Text-to-Image Synthesis
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 |
|---|---|---|---|
| redesign generator: StyleGAN2 backbone, residual convs, stronger conditioning (ablation) | StyleGAN-XL text baseline · Zero-shot MS COCO | FID and CLIP score: ~10% improvement in FID and CLIP score | Helped |
| replace StyleGAN-XL discriminator with ViT-S DINO feature network and five minimalist heads (ablation) | StyleGAN-T · Zero-shot MS COCO | FID and CLIP score: further ~40% improvement | Helped |
| add CLIP guidance loss during training (weight 0.2) (ablation) | StyleGAN-T · Zero-shot MS COCO | FID and CLIP score: further ~20% improvement | Helped |
| fine-tune the text encoder in a secondary phase with frozen generator and CLIP guidance weight 50 (ablation) | StyleGAN-T · Zero-shot MS COCO | CLIP score | Helped |
| use explicit truncation at inference (ablation) | StyleGAN-T · Zero-shot MS COCO | CLIP score: CLIP score pushed to 0.305 | Helped |
| overly strong CLIP guidance weight during primary training (ablation) | StyleGAN-T · Zero-shot MS COCO | FID | Hurt |
| use 1D convolutions on tokens instead of 2D convolutions in discriminator heads (ablation) | StyleGAN-T discriminator · Zero-shot MS COCO | FID | No clear effect |
| drop StyleGAN3 equivariance, use StyleGAN2 backbone (ablation) | StyleGAN-T · Zero-shot MS COCO | FID | Helped |
| increase generator depth with residual convs and layer scale | StyleGAN-T generator · training stability | mode collapse | Helped |
Closest papers by meaning
- Scaling Up GANs for Text-to-Image Synthesis (CVPR 2023 · similarity 0.94)
- Adversarial Text to Continuous Image Generation (CVPR 2024 · similarity 0.89)
- E$^2$GAN: Efficient Training of Efficient GANs for Image-to-Image Translation (ICML 2024 · similarity 0.89)
- You Only Sample Once: Taming One-Step Text-to-Image Synthesis by Self-Cooperative Diffusion GANs (ICLR 2025 · similarity 0.88)
- Exploring Sparse MoE in GANs for Text-conditioned Image Synthesis (CVPR 2025 · similarity 0.88)
- IT3D: Improved Text-to-3D Generation with Explicit View Synthesis (AAAI 2024 · 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.