BK-SDM-v2 Model Card

BK-SDM-{v2-Base, v2-Small, v2-Tiny} are obtained by compressing SD-v2.1-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, GitHub.

Examples with 🤗Diffusers library.

An inference code with the default PNDM scheduler and 50 denoising steps is as follows.

import torch
from diffusers import StableDiffusionPipeline

pipe = StableDiffusionPipeline.from_pretrained("nota-ai/bk-sdm-v2-tiny", 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 and distillation retraining 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+.
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	13.93	35.93	0.3075	0.87B	1.26B
BK-SDM-v2-Base (Ours)	15.85	31.70	0.2868	0.59B	0.98B
BK-SDM-v2-Small (Ours)	16.61	31.73	0.2901	0.49B	0.88B
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	13.05	36.76	0.2958	0.86B	1.04B
BK-SDM-Base (Ours)	15.76	33.79	0.2878	0.58B	0.76B
BK-SDM-Base-2M (Ours)	14.81	34.17	0.2883	0.58B	0.76B
BK-SDM-Small (Ours)	16.98	31.68	0.2677	0.49B	0.66B
BK-SDM-Small-2M (Ours)	17.05	33.10	0.2734	0.49B	0.66B
BK-SDM-Tiny (Ours)	17.12	30.09	0.2653	0.33B	0.50B
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).

Uses

Please follow the usage guidelines of Stable Diffusion v1.

Acknowledgments

Microsoft for Startups Founders Hub and Gwangju AICA for generously providing GPU resources.
CompVis, Runway, and Stability AI for the pioneering research on Stable Diffusion.
LAION, Diffusers, PEFT, DreamBooth, Gradio, and Core ML Stable Diffusion for their valuable contributions.

Citation

@article{kim2023architectural,
  title={BK-SDM: A Lightweight, Fast, and Cheap Version of Stable Diffusion},
  author={Kim, Bo-Kyeong and Song, Hyoung-Kyu and Castells, Thibault and Choi, Shinkook},
  journal={arXiv preprint arXiv:2305.15798},
  year={2023},
  url={https://arxiv.org/abs/2305.15798}
}

@article{kim2023bksdm,
  title={BK-SDM: Architecturally Compressed Stable Diffusion for Efficient Text-to-Image Generation},
  author={Kim, Bo-Kyeong and Song, Hyoung-Kyu and Castells, Thibault and Choi, Shinkook},
  journal={ICML Workshop on Efficient Systems for Foundation Models (ES-FoMo)},
  year={2023},
  url={https://openreview.net/forum?id=bOVydU0XKC}
}

This model card is based on the Stable Diffusion v1 model card.

nota-ai
/

bk-sdm-v2-tiny