DuSPiT: New pixel-space diffusion transformer with dual-branch architecture
Pixel-space diffusion just got a major upgrade with DuSPiT's dual-branch design.
A new architecture, DuSPiT (Dual-Branch Sub-Patch Pixel Diffusion Transformer), is introduced for image generation. Unlike prior pixel-space models that map each raw image patch to a single token, DuSPiT separates global reasoning (via a compact base branch) from local detail modeling (via a parallel, high-capacity pixel branch organized into sub-patch groups), with cross-attention linking both branches. Results show richer details, stronger fine-grained structures, and a better quality-efficiency trade-off than previous pixel-space diffusion transformers.
- DuSPiT uses a dual-branch architecture: one compact branch for global reasoning, another sub-patch branch for local detail.
- Operates entirely in pixel space, avoiding information loss from latent compression.
- Achieves better quality–efficiency trade-off compared to previous pixel-space diffusion transformers.
Why It Matters
DuSPiT pushes pixel-space image generation toward higher fidelity, enabling sharper details for creative and scientific tools.