Research map / Image generation / Text-to-image diffusion
PixArt-Sigma: Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation
Junsong Chen, Chongjian GE, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, Zhenguo Li
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 |
|---|---|---|---|
| compress KV tokens in deep layers (14-27) rather than shallow layers (ablation) | PixArt-Sigma · 512px, High-Quality Eval | FID: 9.278 (shallow) -> 8.532 (deep) | Improved |
| use learnable Conv2x2 for KV compression instead of token discarding or pooling (ablation) | PixArt-Sigma · 512px, High-Quality Eval | FID: 9.415 (pooling) -> 8.505 (Conv2x2) | Improved |
| increase KV compression ratio at 1024px (ablation) | PixArt-Sigma · 1024px | train latency: 27.5 -> 17.8 at ratio 9 | Improved |
| increase KV compression ratio at 1024px (ablation) | PixArt-Sigma · 1024px | FID: 5.685 -> 5.712 | No clear effect |
| increase KV compression ratio (ablation) | PixArt-Sigma · CLIP-Score across resolutions | CLIP-Score: 0.277 -> 0.275 | No clear effect |
| add KV compression (ratio 4) at 4K (ablation) | PixArt-Sigma · 4K | train latency: 191 -> 125 | Improved |
| fine-tune from low-resolution model with PE interpolation | PixArt-alpha · 512px, High-Quality Eval | FID: 16.56 -> 9.75 after 1K steps | Improved |
| weak-to-strong fine-tuning from PixArt-alpha with new data, VAE and captions | PixArt-alpha · High-Quality Eval 30K | FID: 8.65 -> 8.23 | Improved |
| use DMD one-step distillation instead of LCM one-step | PixArt-Sigma · 512px | FID: 108.66 -> 13.35 | Improved |
| set DMD timestep to 400 instead of 999 | PixArt + DMD · 512px | quality | Improved |
Later papers that compared with this method
8 later accepted papers report Pixart-sigma as a baseline in their results tables; 8 of them report a better number on most of the results they share with it.
- L2P: Unlocking Latent Potential for Pixel Generation (NeurIPS 2026 · better on 3 of 3 shared results)
- Latent Wavelet Diffusion For Ultra High-Resolution Image Synthesis (ICLR 2026 · better on 1 of 1 shared result)
- RePrompt: Reasoning-Augmented Reprompting for Text-to-Image Generation via Reinforcement Learning (ICLR 2026 · better on 1 of 1 shared result)
- Everything in Its Place: Benchmarking Spatial Intelligence of Text-to-Image Models (ICLR 2026 · better on 1 of 1 shared result)
- EDiT: Efficient Diffusion Transformers with Linear Compressed Attention (ICCV 2025 · better on 3 of 4 shared results)
- LiT: Delving into a Simple Linear Diffusion Transformer for Image Generation (ICCV 2025 · better on 2 of 3 shared results)
- Infinity: Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis (CVPR 2025 · better on 1 of 1 shared result)
- DC-AR: Efficient Masked Autoregressive Image Generation with Deep Compression Hybrid Tokenizer (ICCV 2025 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- PixArt-$\alpha$: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis (ICLR 2024 · similarity 0.93)
- Diffusion-4K: Ultra-High-Resolution Image Synthesis with Latent Diffusion Models (CVPR 2025 · similarity 0.89)
- SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis (ICLR 2024 · similarity 0.88)
- EC-DIT: Scaling Diffusion Transformers with Adaptive Expert-Choice Routing (ICLR 2025 · similarity 0.87)
- Exploring the Deep Fusion of Large Language Models and Diffusion Transformers for Text-to-Image Synthesis (CVPR 2025 · similarity 0.87)
- On the Scalability of Diffusion-based Text-to-Image Generation (CVPR 2024 · similarity 0.87)
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.
