This demo showcases a lightweight Stable Diffusion model (SDM) for general-purpose text-to-image synthesis. Our model [**BK-SDM-Small**](https://huggingface.co/nota-ai/bk-sdm-small) achieves **36% reduced** parameters and latency. This model is bulit with (i) removing several residual and attention blocks from the U-Net of [SDM-v1.4](https://huggingface.co/CompVis/stable-diffusion-v1-4) and (ii) distillation pretraining on only 0.22M LAION pairs (fewer than 0.1% of the full training set). Despite very limited training resources, our model can imitate the original SDM by benefiting from transferred knowledge.

<center>
    <img alt="U-Net architectures and KD-based pretraining" img src="https://huggingface.co/spaces/nota-ai/compressed-stable-diffusion/resolve/91f349ab3b900cbfec20163edd6a312d1e8c8193/docs/fig_model.png" width="65%">
</center>

<br/>


### Notice
 - The model weights are available at BK-SDM-{[Base](https://huggingface.co/nota-ai/bk-sdm-base), [Small](https://huggingface.co/nota-ai/bk-sdm-small), [Tiny](https://huggingface.co/nota-ai/bk-sdm-tiny)} and can be easily used with 🤗 Diffusers. 
 - This research was accepted to
    - [**ICML 2023 Workshop on Efficient Systems for Foundation Models** (ES-FoMo)](https://es-fomo.com/)
    - [**ICCV 2023 Demo Track**](https://iccv2023.thecvf.com/)
 - Please be aware that your prompts are logged, _without_ any personally identifiable information.
 - For different images with the same prompt, please change _Random Seed_ in Advanced Settings (because of using the firstly sampled latent code per seed).
 
### Acknowledgments
 - We thank [Microsoft for Startups Founders Hub](https://www.microsoft.com/en-us/startups) for supporting this research. 
 - Some demo codes were borrowed from the repo of Stability AI ([stabilityai/stable-diffusion](https://huggingface.co/spaces/stabilityai/stable-diffusion)) and AK ([akhaliq/small-stable-diffusion-v0](https://huggingface.co/spaces/akhaliq/small-stable-diffusion-v0)). Thanks!

### Demo Environment
 - Regardless of machine types, our compressed model achieves speedups while preserving visually compelling results.
 - [June/30/2023] **Free CPU-basic** (2 vCPU · 16 GB RAM) — 7~10 min slow inference of the original SDM.
    - Because free CPU resources are dynamically allocated with other demos, it may take much longer, depending on the server situation.
 - [May/31/2023] **NVIDIA T4-small** (4 vCPU · 15 GB RAM · 16GB VRAM) — 5~10 sec inference of the original SDM (for a 512×512 image with 25 denoising steps).