Reproduce “EdgeDiT: Hardware-Aware Diffusion Transformers for Efficient On-Device Image Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sravanth Kodavanti, Manjunath Arveti, Sowmya Vajrala, Srinivas Miriyala, Vikram N R
Samsung Research Institute Bangalore developed EdgeDiT, a hardware-aware optimization framework that enables efficient on-device image generation using Diffusion Transformers. This framework achieves up to 1.65x faster inference on mobile NPUs with a 20-30% reduction in parameters, while also improving generative quality as measured by FID scores.
Key insights
- Hardware-aware architectural modifications, such as targeted block removal and adjustments to FFN expansion ratios and hidden dimensions, can effectively reduce DiT complexity while preserving generative capacity.
- Feature-wise Knowledge Distillation is critical for efficiently pre-training compressed model components, enabling the construction of full candidate architectures without exhaustive retraining.
- Multi-objective Bayesian Optimization, incorporating real-world NPU latency as an objective, is essential for identifying models that offer optimal trade-offs between generative quality and on-device efficiency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.