Research map / Image generation / Autoregressive and latent generation
DiverseDiT: Towards Diverse Representation Learning in Diffusion Transformers
Mengping Yang, Zhiyu Tan, Binglei Li, Xiaomeng Yang, Hesen Chen, Hao 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 |
|---|---|---|---|
| add DiverseDiT (long residuals + diversity loss) | SiT-B · ImageNet 256x256 no CFG, 400K | FID: 36.80 -> 28.05 | Improved |
| add DiverseDiT | REPA-B · ImageNet 256x256 no CFG, 400K | FID: 22.99 -> 17.29 | Improved |
| add DiverseDiT | REPA-XL · ImageNet 256x256 no CFG, 400K | FID: 8.73 -> 8.09 | Improved |
| add DiverseDiT | MeanFlow-XL/2 (one-step) · ImageNet 256x256 one-step, no CFG | FID: 5.94 -> 5.69 | Improved |
| add DiverseDiT instead of DispLoss | MeanFlow-XL/2 (one-step, CFG) · ImageNet 256x256 one-step with CFG | FID: 3.21 -> 2.99 | Improved |
| add representation diversity loss (ablation) | SiT-B + DiverseDiT · ImageNet 256x256 no CFG, 400K | FID: 32.77 -> 28.05 | Improved |
| add long residual connections (ablation) | SiT-B + DiverseDiT · ImageNet 256x256 no CFG, 400K | FID: 33.72 -> 28.05 | Improved |
| add long residual connections (ablation) | REPA-B + DiverseDiT · ImageNet 256x256 no CFG, 400K | FID: 18.18 -> 17.29 | Improved |
| combine orthogonality, mutual-information and dispersion losses instead of one (ablation) | REPA-B + DiverseDiT · ImageNet 256x256 no CFG, 400K | FID: 18.97 -> 17.29 (vs only Lorth) | Improved |
| use adaptive loss weight range [0.1, 0.5] instead of wider ranges (ablation) | SiT-B + DiverseDiT · ImageNet 256x256 no CFG, 400K | FID: 31.85 -> 28.05 | Improved |
| use constant diversity-loss weight instead of adaptive (ablation) | SiT-B + DiverseDiT · ImageNet 256x256 no CFG | FID: 28.05 -> diverge | Hurt |
| add DispLoss and SRA on top of DiverseDiT | SiT-B · ImageNet 256x256 no CFG, 400K | FID: 28.05 -> 24.98 -> 21.95 | Improved |
Closest papers by meaning
- Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think (ICLR 2025 · similarity 0.90)
- Why DiT Models Underperform as Representation Learners without Long Skip Connections (NeurIPS 2026 · similarity 0.89)
- UDT: Reconciling U-Nets and Diffusion Transformers with Data-Adaptive Token Reduction (NeurIPS 2026 · similarity 0.89)
- BlockDance: Reuse Structurally Similar Spatio-Temporal Features to Accelerate Diffusion Transformers (CVPR 2025 · similarity 0.89)
- FasterDiT: Towards Faster Diffusion Transformers Training without Architecture Modification (NeurIPS 2024 · similarity 0.88)
- Unleashing Diffusion Transformers for Visual Correspondence by Modulating Massive Activations (NeurIPS 2025 · 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.
