---
license: creativeml-openrail-m
tags:
- stable-diffusion
- stable-diffusion-diffusers
- text-to-image
datasets:
- ChristophSchuhmann/improved_aesthetics_6.5plus
library_name: diffusers
pipeline_tag: text-to-image
extra_gated_prompt: >-
This model is open access and available to all, with a CreativeML OpenRAIL-M
license further specifying rights and usage.
The CreativeML OpenRAIL License specifies:
1. You can't use the model to deliberately produce nor share illegal or
harmful outputs or content
2. The authors claim no rights on the outputs you generate, you are free to
use them and are accountable for their use which must not go against the
provisions set in the license
3. You may re-distribute the weights and use the model commercially and/or as
a service. If you do, please be aware you have to include the same use
restrictions as the ones in the license and share a copy of the CreativeML
OpenRAIL-M to all your users (please read the license entirely and carefully)
Please read the full license carefully here:
https://huggingface.co/spaces/CompVis/stable-diffusion-license
extra_gated_heading: Please read the LICENSE to access this model
---
# BK-SDM-v2 Model Card
BK-SDM-{[**v2-Base**](https://huggingface.co/nota-ai/bk-sdm-v2-base), [**v2-Small**](https://huggingface.co/nota-ai/bk-sdm-v2-small), [**v2-Tiny**](https://huggingface.co/nota-ai/bk-sdm-v2-tiny)} are obtained by compressing [SD-v2.1-base](https://huggingface.co/stabilityai/stable-diffusion-2-1-base).
- Block-removed Knowledge-distilled Stable Diffusion Models (BK-SDMs) are developed for efficient text-to-image (T2I) synthesis:
- Certain residual & attention blocks are eliminated from the U-Net of SD.
- Despite the use of very limited data, distillation retraining remains surprisingly effective.
- Resources for more information: [Paper](https://arxiv.org/abs/2305.15798), [GitHub](https://github.com/Nota-NetsPresso/BK-SDM).
## Examples with 🤗[Diffusers library](https://github.com/huggingface/diffusers).
An inference code with the default PNDM scheduler and 50 denoising steps is as follows.
```python
import torch
from diffusers import StableDiffusionPipeline
pipe = StableDiffusionPipeline.from_pretrained("nota-ai/bk-sdm-v2-base", torch_dtype=torch.float16)
pipe = pipe.to("cuda")
prompt = "a black vase holding a bouquet of roses"
image = pipe(prompt).images[0]
image.save("example.png")
```
## Compression Method
Based on the [U-Net architecture](https://huggingface.co/nota-ai/bk-sdm-base#u-net-architecture) and [distillation retraining](https://huggingface.co/nota-ai/bk-sdm-base#distillation-pretraining) of BK-SDM, a reduced batch size (from 256 to 128) is used in BK-SDM-v2 for faster training speeds.
- **Training Data**: 212,776 image-text pairs (i.e., 0.22M pairs) from [LAION-Aesthetics V2 6.5+](https://laion.ai/blog/laion-aesthetics/).
- **Hardware:** A single NVIDIA A100 80GB GPU
- **Gradient Accumulations**: 4
- **Batch:** 128 (=4×32)
- **Optimizer:** AdamW
- **Learning Rate:** a constant learning rate of 5e-5 for 50K-iteration retraining
## Experimental Results
The following table shows the zero-shot results on 30K samples from the MS-COCO validation split. After generating 512×512 images with the PNDM scheduler and 25 denoising steps, we downsampled them to 256×256 for evaluating generation scores.
- Our models were drawn at the 50K-th training iteration.
#### Compression of SD-v2.1-base
| Model | FID↓ | IS↑ | CLIP Score↑
(ViT-g/14) | # Params,
U-Net | # Params,
Whole SDM |
|---|:---:|:---:|:---:|:---:|:---:|
| [Stable Diffusion v2.1-base](https://huggingface.co/stabilityai/stable-diffusion-2-1-base) | 13.93 | 35.93 | 0.3075 | 0.87B | 1.26B |
| [BK-SDM-v2-Base](https://huggingface.co/nota-ai/bk-sdm-v2-base) (Ours) | 15.85 | 31.70 | 0.2868 | 0.59B | 0.98B |
| [BK-SDM-v2-Small](https://huggingface.co/nota-ai/bk-sdm-v2-small) (Ours) | 16.61 | 31.73 | 0.2901 | 0.49B | 0.88B |
| [BK-SDM-v2-Tiny](https://huggingface.co/nota-ai/bk-sdm-v2-tiny) (Ours) | 15.68 | 31.64 | 0.2897 | 0.33B | 0.72B |
#### Compression of SD-v1.4
| Model | FID↓ | IS↑ | CLIP Score↑
(ViT-g/14) | # Params,
U-Net | # Params,
Whole SDM |
|---|:---:|:---:|:---:|:---:|:---:|
| [Stable Diffusion v1.4](https://huggingface.co/CompVis/stable-diffusion-v1-4) | 13.05 | 36.76 | 0.2958 | 0.86B | 1.04B |
| [BK-SDM-Base](https://huggingface.co/nota-ai/bk-sdm-base) (Ours) | 15.76 | 33.79 | 0.2878 | 0.58B | 0.76B |
| [BK-SDM-Base-2M](https://huggingface.co/nota-ai/bk-sdm-base-2m) (Ours) | 14.81 | 34.17 | 0.2883 | 0.58B | 0.76B |
| [BK-SDM-Small](https://huggingface.co/nota-ai/bk-sdm-small) (Ours) | 16.98 | 31.68 | 0.2677 | 0.49B | 0.66B |
| [BK-SDM-Small-2M](https://huggingface.co/nota-ai/bk-sdm-small-2m) (Ours) | 17.05 | 33.10 | 0.2734 | 0.49B | 0.66B |
| [BK-SDM-Tiny](https://huggingface.co/nota-ai/bk-sdm-tiny) (Ours) | 17.12 | 30.09 | 0.2653 | 0.33B | 0.50B |
| [BK-SDM-Tiny-2M](https://huggingface.co/nota-ai/bk-sdm-tiny-2m) (Ours) | 17.53 | 31.32 | 0.2690 | 0.33B | 0.50B |
#### Visual Analysis: Image Areas Affected By Each Word
KD enables our models to mimic the SDM, yielding similar per-word attribution maps. The model without KD behaves differently, causing dissimilar maps and inaccurate generation (e.g., two sheep and unusual bird shapes).