This demo showcases a lightweight Stable Diffusion model (SDM) for general-purpose text-to-image synthesis. Our model **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 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/theme/resolve/3bb3eed8b911d0baf306767bb9548bf732052c53/docs/compressed_stable_diffusion/fig_model.png" width="65%">
</center>

<br/>


### Notice
 - This research is accepted to [**ICCV 2023 Demo Track**](https://iccv2023.thecvf.com/) — title: Architecturally Compressed Stable Diffusion for Efficient Text-to-Image Generation.
 - Please be aware that your prompts are logged (_without_ any personally identifiable information).
 - To generate different images with the same prompt, please change _Random Seed_ in Advanced Settings (because this demo only uses the firstly sampled latent code for each seed).
 - Many parts of the demo codes were borrowed from [stabilityai/stable-diffusion](https://huggingface.co/spaces/stabilityai/stable-diffusion) and [akhaliq/small-stable-diffusion-v0](https://huggingface.co/spaces/akhaliq/small-stable-diffusion-v0). Thanks, Stability AI ([@stabilityai](https://huggingface.co/stabilityai)) and AK ([@akhaliq](https://huggingface.co/akhaliq))!

### Updates
 - [May/31/2023] The demo is running on **T4-small** (4 vCPU · 15 GB RAM · 16GB VRAM). It takes 5~10 seconds for the original model to generate a 512×512 image with 25 denoising steps. Our compressed model accelerates inference speed while preserving visually compelling results.