Research map / Image generation / Autoregressive and latent generation
PixelDiT: Pixel Diffusion Transformers for Image Generation
Yongsheng Yu, Wei Xiong, Weili Nie, Yichen Sheng, Shiqiu Liu, Jiebo Luo
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 |
|---|---|---|---|
| add RoPE and RMSNorm (ablation) | Vanilla DiT/16 · ImageNet 256x256 ablation (80 epochs) | gFID: 9.84 -> 8.53 | Improved |
| add dual-level pathway with pixel token compaction (ablation) | Vanilla DiT/16 · ImageNet 256x256 ablation (80 epochs) | gFID: 8.53 -> 3.50 | Improved |
| add pixel-wise AdaLN (ablation) | PixelDiT-XL · ImageNet 256x256 ablation (80 epochs) | gFID: 3.50 -> 2.36 | Improved |
| add pixel token compaction to dual-level design (ablation) | PixelDiT-XL · ImageNet 256x256 compute | GFLOPs: 82247 -> 311 | Improved |
| add self-attention in the pixel-level pathway (ablation) | PixelDiT-XL · ImageNet 256x256 ablation (80 epochs) | gFID: 2.56 -> 2.36 | Improved |
| add representation alignment (REPA) (ablation) | PixelDiT-XL · ImageNet 256x256 ablation (80 epochs) | gFID: 6.58 -> 2.36 | Improved |
| use compaction to 1x sequence length instead of 4x (ablation) | PixelDiT · ImageNet 256x256 ablation (300K iterations) | gFID: 2.43 -> 2.34 | Improved |
| use patch size 16 instead of 32 (ablation) | PixelDiT · ImageNet 512x512 ablation (120 epochs) | FID: 3.78 -> 2.23 | Improved |
| use patch size 16 instead of 32 (ablation) | PixelDiT-B · ImageNet 256x256 ablation (200K iterations) | gFID: 48.5 -> 15.1 | Improved |
| use patch size 8 instead of 16 (ablation) | PixelDiT-L · ImageNet 256x256 ablation (300K iterations) | gFID: 2.72 -> 2.15 | Improved |
| use patch size 8 instead of 16 (ablation) | PixelDiT-XL · ImageNet 256x256 ablation | gFID | No clear effect |
| scale model from B/16 to XL/16 (ablation) | PixelDiT · ImageNet 256x256 ablation (200K iterations) | gFID: 15.1 -> 2.95 | Improved |
Later papers that compared with this method
2 later accepted papers report PixelDiT as a baseline in their results tables; 2 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 2 of 2 shared results)
- Hi-DiT: Hybrid Latent-Pixel Diffusion Transformer for Image Generation (ECCV 2026 · 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
- PixelDiT2: Representation-Grounded Pixel Diffusion Transformers (NeurIPS 2026 · similarity 0.94)
- Simpler Diffusion: 1.5 FID on ImageNet512 with Pixel-space Diffusion (CVPR 2025 · similarity 0.91)
- PixelART: Image-to-Layer Decomposition without Latents or Text-to-Image Pretraining (NeurIPS 2026 · similarity 0.91)
- Scalable High-Resolution Pixel-Space Image Synthesis with Hourglass Diffusion Transformers (ICML 2024 · similarity 0.89)
- L2P: Unlocking Latent Potential for Pixel Generation (NeurIPS 2026 · similarity 0.89)
- PiD: Fast and High-Resolution Latent Decoding with Pixel Diffusion (NeurIPS 2026 · 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.
