AI Briefing5.1x Faster Stable Diffusion on Intel CPUs
Key point
This presents an optimization method using NNCF and Optimum that boosts Stable Diffusion inference speed on Intel CPUs by 5.1x.
Details
The UNet model, the core of Stable Diffusion, requires a very large amount of computation, which results in very slow generation speed when running on CPU alone without a GPU.
The existing 8-bit quantization (Post-training quantization) method causes a significant drop in accuracy for complex models like Stable Diffusion. This is because the model has low redundancy and contains a vast amount of information.
To address this, Hugging Face proposes the following workflow:
- Applying Quantization-Aware Training (QAT) via NNCF
- Using Knowledge Distillation to maintain accuracy during the optimization process
- Combining with Token Merging technology
Applying this method achieved a 5.1x inference speedup and a 4x reduction in model size compared to PyTorch.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
2026 AI Inference Hardware Innovations: New Architectures for Prefill/Decode Disaggregation and Solving Memory Bottlenecks
Pinterest Reduces Costs by Up to 30% with SSD Serving and Quantization in Manas Embedding Search Platform