Research map / Architectures and efficiency / Spiking networks
Trajectory-Aware Spiking DiTs Conversion via Membrane Potential Error-Feedback
Haoran Fang, Tianxing Man, Xingchen Li, Wanli Shi, Jinjie Fang, Bin Gu
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 Hadamard rotation to naive conversion (ablation) | Spiking DiT conversion · DiT-XL-2-256 (T=4, n=6) | FID: 198.03 -> 18.47 | Improved |
| add MPEF (carry residual membrane potential across denoising steps) to naive conversion (ablation) | Spiking DiT conversion · DiT-XL-2-256 (T=4, n=6) | FID: 198.03 -> 7.01 | Improved |
| add MPEF on top of Hadamard rotation (ablation) | Spiking DiT conversion · DiT-XL-2-256 (T=4, n=6) | FID: 18.47 -> 6.52 | Improved |
| add Hadamard rotation on top of MPEF (ablation) | Spiking DiT conversion · DiT-XL-2-256 (T=4, n=6) | FID: 7.01 -> 6.52 | Improved |
| increase multi-threshold count n from 6 to 8 (ablation) | Spiking DiT · ImageNet 256x256, 50 steps, cfg=1.0 | FID: 19.29 -> 11.72 | Improved |
| increase multi-threshold count n from 6 to 8 (ablation) | Spiking DiT · ImageNet 256x256, 50 steps, cfg=1.5 | FID: 5.86 -> 5.59 | Improved |
| increase multi-threshold count n from 8 to 10 (ablation) | Spiking DiT · ImageNet 512x512, 50 steps, cfg=1.5 | FID: 5.91 -> 5.93 | No clear effect |
| spiking conversion replaces MAC with AC operations | Spiking DiT · DiT-XL-2-256 operations | MACs (G): 118.69 -> 0.91 | Improved |
| Spiking DiT vs W4A8 PTQ baselines at low bit-width | Spiking DiT (n=6) · ImageNet 256x256, 50 steps, cfg=1.5 | FID: Q-DiT W4A8 6.54 -> 5.86 | Improved |
| use n=6 spiking at cfg=1.0 vs Q-DiT W8A8 is worse than full precision | Spiking DiT (n=8) · ImageNet 256x256, 50 steps, cfg=1.0 | FID: FP 10.99 vs ours 11.72 | No clear effect |
Closest papers by meaning
- Adaptive Calibration: A Unified Conversion Framework of Spiking Neural Networks (AAAI 2025 · similarity 0.89)
- Rethinking the Membrane Dynamics and Optimization Objectives of Spiking Neural Networks (NeurIPS 2024 · similarity 0.89)
- STEP: A Unified Spiking Transformer Evaluation Platform for Fair and Reproducible Benchmarking (NeurIPS 2025 · similarity 0.88)
- Efficient ANN-SNN Conversion with Error Compensation Learning (ICML 2025 · similarity 0.88)
- Faster and Stronger: When ANN-SNN Conversion Meets Parallel Spiking Calculation (ICML 2025 · similarity 0.88)
- Spike-driven Transformer V2: Meta Spiking Neural Network Architecture Inspiring the Design of Next-generation Neuromorphic Chips (ICLR 2024 · similarity 0.87)
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.
