README.md

---
license: apache-2.0
datasets:
- poloclub/diffusiondb
base_model:
- stabilityai/stable-diffusion-2-1-base
pipeline_tag: text-to-image
library_name: diffusers
---
# AMD Nitro-1


![image/jpeg](https://cdn-uploads.huggingface.co/production/uploads/6355aded9c72a7e742f341a4/AsUvS7acUDLZhKOMRSH37.jpeg)

## Introduction
Nitro-1 is a series of efficient text-to-image generation models that are distilled from popular diffusion models on AMD Instinct™ GPUs. The release consists of:

* [Nitro-1-SD](https://huggingface.co/amd/SD2.1-Nitro): a UNet-based one-step model distilled from [Stable Diffusion 2.1](https://huggingface.co/stabilityai/stable-diffusion-2-1-base).
* [Nitro-1-PixArt](https://huggingface.co/amd/PixArt-Sigma-Nitro): a high resolution transformer-based one-step model distilled from [PixArt-Sigma](https://pixart-alpha.github.io/PixArt-sigma-project/).

⚡️ [Open-source code](https://github.com/AMD-AIG-AIMA/AMD-Diffusion-Distillation)! The models are based on our re-implementation of [Latent Adversarial Diffusion Distillation](https://arxiv.org/abs/2403.12015), the method used to build the popular Stable Diffusion 3 Turbo model. Since the original authors didn't provide training code, we release our re-implementation to help advance further research in the field.



## Details

* **Model architecture**: Nitro-1-SD has the same architecture as Stable Diffusion 2.1 and is compatible with the diffusers pipeline.
* **Inference steps**: This model is distilled to perform inference in just a single step. However, the training code also supports distilling a model for 2, 4 or 8 steps.
* **Hardware**: We use a single node consisting of 4 AMD Instinct™ MI250 GPUs for distilling Nitro-1-SD.
* **Dataset**: We use 1M prompts from [DiffusionDB](https://huggingface.co/datasets/poloclub/diffusiondb) and generate the corresponding images from the base Stable Diffusion 2.1 model.
* **Training cost**: The distillation process achieves reasonable results in less than 2 days on a single node.



## Quickstart

```python
from diffusers import DDPMScheduler, DiffusionPipeline
import torch

scheduler = DDPMScheduler.from_pretrained("stabilityai/stable-diffusion-2-1-base", subfolder="scheduler")
pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1-base", scheduler=scheduler)

ckpt_path = '<path to distilled checkpoint>'
unet_state_dict = torch.load(ckpt_path)
pipe.unet.load_state_dict(unet_state_dict)
pipe = pipe.to("cuda")

image = pipe(prompt='a photo of a cat',
             num_inference_steps=1,
             guidance_scale=0,
             timesteps=[999]).images[0]
```

For more details on training and evaluation please visit the [GitHub repo](https://github.com/AMD-AIG-AIMA/AMD-Diffusion-Distillation).



## Results


Compared to the [Stable Diffusion 2.1 base model](https://huggingface.co/stabilityai/stable-diffusion-2-1-base), we achieve 95.9% reduction in FLOPs at the cost of just 2.5% lower CLIP score and 2.2% higher FID.

| Model    | FID &darr; | CLIP &uarr; |FLOPs| Latency on AMD Instinct MI250 (sec)
| :---: | :---: | :---: | :---: | :---:
| Stable Diffusion 2.1 base, 50 steps (cfg=7.5) | 25.47   | 0.3286 |83.04 | 4.94
| **Nitro-1-SD**, 1 step | 26.04     | 0.3204|3.36 | 0.18



## License
Copyright (c) 2018-2024 Advanced Micro Devices, Inc. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
    http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.