NTU's Persistence Forcing hits FID 1.63 on ImageNet 256 by heterogeneous refinement in pixel-space DiTs
NanyangTechnologicalUniversity · hf · 2026-09-30
Nanyang Technological University introduces Persistence Forcing (PerF), exploiting emergent feature specialization in pixel-space diffusion Transformers.
- Observation: pixel-space DiTs refine hidden representations uniformly across depth, but natural images need richer representations for local textures than for global structure.
- Method: heterogeneous refinement assigns different feature groups distinct refinement budgets, yielding ordered specialization into "persistent" features (global structure) and "active" features (local high-frequency detail).
- During sampling, persistent features continuously condition active ones, promoting coherent global structure and complementing classifier-free guidance.
- Results: PerF-L achieves FID 1.91 on ImageNet 256×256 with half the parameters of JiT-H (1.86); PerF-H reaches 1.63 and 1.76 on 256×256 and 512×512.
More from Research
- Triangle Splatting SLAM: Imperial College's ECCV 2026 dense RGB-D SLAM with on-the-fly mesh extraction — rsasaki0109 · 2026-09-30
- Manifold opens early access: robotics eval platform runs thousands of GPU-parallel rollouts in 30 mins — paigeinsf · 2026-09-30
- 1,000 AI agents discover new CRISPR-like system in virus DNA within 24 hours — CurieuxExplorer · 2026-09-30
- Explaining just 5% of token positions retains nearly all audit success across 4.7M explanations — aisilab · 2026-09-30
- IBM's Q&D trains proactive agents to ask better questions, beating a 15x larger model — ibm · 2026-09-30
- Researcher predicts AI labs will soon pivot from math conjectures to materials and drug discovery — tak3sh8 · 2026-09-30