Research map / Image generation / Text-to-image diffusion
Unveiling the Secret of AdaLN-Zero in Diffusion Transformer
Jie Zhu, Mingyu Ding, Boqiang Duan, Leye Wang, Jingdong Wang
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 |
|---|---|---|---|
| initialize W_alpha, W_gamma, W_beta with Gaussian std 1e-3 instead of zero (ablation) | DiT-XL/2 · ImageNet 256 (50K steps) | FID: 78.99 -> 76.21 | Improved |
| use too large Gaussian std 5e-2 (ablation) | adaLN-Gaussian · ImageNet 256 (50K steps) | FID: 78.99 -> 84.37 | Hurt |
| use too small Gaussian std 5e-4 (ablation) | adaLN-Gaussian · ImageNet 256 (50K steps) | FID: 78.99 -> 80.68 | Hurt |
| Gaussian init (1e-3) only for W_alpha, W_gamma and W_beta left zero (ablation) | adaLN-Zero · ImageNet 256 (50K steps) | FID: 78.99 -> 78.62 | No clear effect |
| also Gaussian-initialize W_gamma and W_beta (ablation) | adaLN-Gaussian W_alpha only · ImageNet 256 (50K steps) | FID: 78.62 -> 76.21 | Improved |
| replace zero init with adaLN-Gaussian at 400K steps | DiT-XL/2 · ImageNet 256, CFG 1 | FID: 20.02 -> 17.86 | Improved |
| replace zero init with adaLN-Gaussian at 400K steps with CFG 1.5 | DiT-XL/2 · ImageNet 256, CFG 1.5 | FID: 6.15 -> 5.28 | Improved |
| adaLN-Gaussian at very long training 7000K steps | DiT-XL/2 · ImageNet 256, CFG 1.5 | FID: 2.27 -> 2.21 | Improved |
| adaLN-Gaussian on smaller DiT-B/2 | DiT-B/2 · ImageNet 256, 400K | FID: 42.72 -> 42.55 | No clear effect |
| adaLN-Gaussian on DiT-L/2 | DiT-L/2 · ImageNet 256, 400K | FID: 24.40 -> 23.05 | Improved |
| adaLN-Gaussian on LlamaVision-XL/2 | LlamaVision-XL/2 · ImageNet 256, 400K | FID: 21.66 -> 20.26 | Improved |
| adaLN-Gaussian on text-to-image | DiT-XL/2 · COCO FID-30K | FID: 71.41 -> 65.51 | Improved |
Closest papers by meaning
- Accelerating Diffusion Transformer via Increment-Calibrated Caching with Channel-Aware Singular Value Decomposition (CVPR 2025 · similarity 0.89)
- Effective Diffusion Transformer Architecture for Image Super-Resolution (AAAI 2025 · similarity 0.89)
- SD-DiT: Unleashing the Power of Self-supervised Discrimination in Diffusion Transformer (CVPR 2024 · similarity 0.89)
- Dynamic Diffusion Transformer (ICLR 2025 · similarity 0.89)
- FasterDiT: Towards Faster Diffusion Transformers Training without Architecture Modification (NeurIPS 2024 · similarity 0.88)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 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.
