Research map / Image generation / Autoregressive and latent generation
Multi-Patch Global-to-Local Transformer Architecture For Efficient Flow Matching and Diffusion Model
Quan Dao, Dimitris Metaxas
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 |
|---|---|---|---|
| replace isotropic DiT with multi-patch global-to-local design (k=6) (ablation) | DiT-B/2 · ImageNet 256, B, 80 epochs | GFLOPs: 23 -> 16.6 | Improved |
| replace isotropic DiT with multi-patch design (k=6) (ablation) | DiT-B/2 + embeddings · ImageNet 256, B, 80 epochs | FID: 24.52 -> 24.74 | No clear effect |
| share AdaIN module across all blocks (ablation) | DiT-B/2 · ImageNet 256, B, 80 epochs | FID: 34.84 -> 35.31 (Params 130.0M -> 90.3M) | Hurt |
| use 16 class tokens instead of single class token (ablation) | DiT-B/2 with shared AdaIN · ImageNet 256, B, 80 epochs | FID: 35.31 -> 28.56 | Improved |
| use FNO time embedding instead of sinusoidal linear embedding (ablation) | DiT-B/2 with multi-token class · ImageNet 256, B, 80 epochs | FID: 28.56 -> 24.52 | Improved |
| increase number of layers in time embedding (ablation) | time embedding · ImageNet 256, B | FID | Hurt |
| replace 1D grid time feature with sinusoidal time feature in FNO embedding (ablation) | FNO time embedding · ImageNet 256 | FID: model does not converge | Hurt |
| increase class tokens from 16 to 32 (ablation) | multi-token class embedding · ImageNet 256, B | FID | No clear effect |
| replace ConvTranspose with Linear layer in upsample block (ablation) | MPDiT · ImageNet 256, B | FID | Improved |
| use MLP or Conv refinement after upsampling instead of one Linear layer (ablation) | MPDiT upsample block · ImageNet 256, B | FID | Hurt |
| use three-level patch hierarchy {8,4,2} instead of two-level {4,2} (ablation) | MPDiT-XL · ImageNet 512 | FID: FID 2.47 -> 3.13 with cfg; GFLOPs 228.4 -> 138.2 | Mixed |
| use MPDiT-XL instead of SiT-XL/2 | MPDiT-XL · ImageNet 256, 80 epochs | FID: 18.04 -> 9.92, GFLOPs 118.66 -> 59.30 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DiCo-XL — DiCo: Revitalizing ConvNets for Scalable and Efficient Diffusion Modeling (NeurIPS 2025) · better on 2 of 2 shared results
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
- Expert Race: A Flexible Routing Strategy for Scaling Diffusion Transformer with Mixture of Experts (ICML 2025 · similarity 0.89)
- Rethinking Cross-Layer Information Routing in Diffusion Transformer (NeurIPS 2026 · similarity 0.89)
- LaVin-DiT: Large Vision Diffusion Transformer (CVPR 2025 · similarity 0.88)
- Layer- and Timestep-Adaptive Differentiable Token Compression Ratios for Efficient Diffusion Transformers (CVPR 2025 · similarity 0.88)
- Exploring Diffusion Transformer Designs via Grafting (NeurIPS 2025 · 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.
