File size: 14,785 Bytes

---
license: apache-2.0
pipeline_tag: mask-generation
library_name: mlx
tags:
- mlx
- moonbox
- efficientvit
- efficientvit-sam
- segment-anything
- image-segmentation
---

# EfficientViT-SAM MLX

MLX builds of [EfficientViT-SAM](https://github.com/mit-han-lab/efficientvit), made by Moonbox for Apple Silicon Macs. Each precision is a folder; keep a folder's
files together. Only the parts listed below are converted: the rest of the pipeline comes from the original
repository unless the section below says otherwise.

## Files

The last column is how far each build is from the original PyTorch model on the same inputs (see the section below
for what was measured).

| Precision | File | Size | sha256 | Checked against the original |
|---|---|---|---|---|
| bf16 | [`bf16/config.json`](https://huggingface.co/moonbox-ai/efficientvit-MLX/blob/main/bf16/config.json) | 0.3 kB | `9e98792690f7` | below threshold (experimental): mask IoU 0.714 mean / 0.091 worst |
| bf16 | [`bf16/mlx_model.py`](https://huggingface.co/moonbox-ai/efficientvit-MLX/blob/main/bf16/mlx_model.py) | 25.4 kB | `42cea84fb312` | below threshold (experimental): mask IoU 0.714 mean / 0.091 worst |
| bf16 | [`bf16/model.safetensors`](https://huggingface.co/moonbox-ai/efficientvit-MLX/blob/main/bf16/model.safetensors) | 122.7 MB | `e7acc4ef2253` | below threshold (experimental): mask IoU 0.714 mean / 0.091 worst |
| fp16 | [`fp16/config.json`](https://huggingface.co/moonbox-ai/efficientvit-MLX/blob/main/fp16/config.json) | 0.3 kB | `9b286f991620` | below threshold (experimental): mask IoU 0.942 mean / 0.611 worst (float32 port: 0.9996 / 0.997) |
| fp16 | [`fp16/mlx_model.py`](https://huggingface.co/moonbox-ai/efficientvit-MLX/blob/main/fp16/mlx_model.py) | 25.4 kB | `42cea84fb312` | below threshold (experimental): mask IoU 0.942 mean / 0.611 worst (float32 port: 0.9996 / 0.997) |
| fp16 | [`fp16/model.safetensors`](https://huggingface.co/moonbox-ai/efficientvit-MLX/blob/main/fp16/model.safetensors) | 122.7 MB | `f4b4b159c9f7` | below threshold (experimental): mask IoU 0.942 mean / 0.611 worst (float32 port: 0.9996 / 0.997) |

## How to run

EfficientViT-SAM **L2** (EfficientViT-L2 image encoder at 512 px + SAM's prompt encoder and mask decoder) ported to
MLX. Each folder holds `model.safetensors`, `config.json` and `mlx_model.py`, the port (only `mlx`, `numpy` and
`pillow` needed), with EfficientViTSamPredictor's interface:

```python
import numpy as np, mlx_model
from PIL import Image
model = mlx_model.load("fp16")
rgb = np.array(Image.open("in.jpg").convert("RGB"))
emb = mlx_model.embed(model, rgb)
masks, scores, low_res, logits = mlx_model.predict(model, emb, points=[[320, 240]], labels=[1], multimask=True)
```

**Both folders are experimental: this model needs float32.** Thirty of its 1x1 convolutions are followed by a
BatchNorm with a running variance down to 5.5e-7 (scales up to 600): those channels carry their signal in the 4th-5th
significant digit of the convolution output, which 16-bit weights or arithmetic can't hold. The port itself is exact:
run in float32 (with `MLX_ENABLE_TF32=0`, else MLX's float32 matmuls use TF32) it matches PyTorch to mask IoU 0.9996.
There are no 8bit / 6bit / 4bit folders: quantized, the model gives no usable masks.

| Precision | Size | Mask IoU vs PyTorch fp32 (mean / worst) | Status |
|---|---|---|---|
| `fp16/` | 122.7 MB | 0.942 / 0.611 | **below threshold (0.95), experimental** |
| `bf16/` | 122.7 MB | 0.714 / 0.091 | **below threshold, experimental** |

To run in float32, load either folder and cast the parameters up (`model.update(mlx.utils.tree_map(lambda a:
a.astype(mx.float32), model.parameters()))`): fp16 weights in float32 give 0.966 mean / 0.62 worst, still short of
the original's float32 weights. About 55 ms per image (encoder + 2 prompts) on an Apple M5.

## How these were made

From [EfficientViT-SAM](https://github.com/mit-han-lab/efficientvit) at revision `de7d7733cc`: the original weights loaded in MLX and stored as bf16 or fp16, or
quantized with MLX's affine quantization (`8bit`, `6bit`, `4bit`, group size 64; layers MLX can't quantize, such as
convolutions and norms, stay bf16). Every build was checked against the original before upload; the `sha256` column
above is what the files must hash to.

## License

apache-2.0. These files are a derivative of the original model and carry its license: it is copied in this
repository, and it governs your use of these files. The original model card follows unchanged.

---

# Original model card

# Efficient Vision Foundation Models for High-Resolution Generation and Perception

[![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/deep-compression-autoencoder-for-efficient/image-generation-on-imagenet-512x512)](https://paperswithcode.com/sota/image-generation-on-imagenet-512x512?p=deep-compression-autoencoder-for-efficient)

## News
- (🔥 New) [2025/09/05] We will no longer maintain this codebase. All future updates and announcements will be made on [DC-Gen](https://github.com/dc-ai-projects/DC-Gen).
- (🔥 New) [2025/01/24] We released DC-AE-SANA-1.1: [doc](https://github.com/mit-han-lab/efficientvit/blob/master/assets/docs/dc_ae_sana_1.1.md). 
- (🔥 New) [2025/01/23] DC-AE and SANA are accepted by ICLR 2025.
- (🔥 New) [2025/01/14] We released **DC-AE+USiT models**: [model](https://huggingface.co/collections/mit-han-lab/dc-ae-diffusion-670dbb8d6b6914cf24c1a49d), [training](https://github.com/mit-han-lab/efficientvit/blob/master/applications/dc_ae/README.md#dc-ae--usit). Using the default training settings and sampling strategy, DC-AE+USiT-2B achieves 1.72 FID on ImageNet 512x512, surpassing the SOTA diffusion model EDM2-XXL and SOTA auto-regressive image generative models (MAGVIT-v2 and MAR-L).

______________________________________________________________________

- (🔥 New) [2024/12/24] **diffusers** supports DC-AE models. All [DC-AE models in diffusers safetensors](https://huggingface.co/collections/mit-han-lab/dc-ae-670085b9400ad7197bb1009b) are released. [Usage](https://github.com/mit-han-lab/efficientvit/tree/master/applications/dc_ae#deep-compression-autoencoder-diffusers).
- [2024/10/21] DC-AE and EfficientViT block are used in our latest text-to-image diffusion model SANA! Check the [project page](https://nvlabs.github.io/Sana/) for more details.
- [2024/10/15] We released **Deep Compression Autoencoder (DC-AE)**: [link](#deep-compression-autoencoder-for-efficient-high-resolution-diffusion-models-paper-readme)!
- [2024/07/10] EfficientViT is used as the backbone in [Grounding DINO 1.5 Edge](https://arxiv.org/pdf/2405.10300) for efficient open-set object detection.
- [2024/07/10] EfficientViT-SAM is used in [MedficientSAM](https://github.com/hieplpvip/medficientsam), the 1st place model in [CVPR 2024 Segment Anything In Medical Images On Laptop Challenge](https://www.codabench.org/competitions/1847/).
- [2024/04/06] EfficientViT-SAM is accepted by [eLVM@CVPR'24](https://sites.google.com/view/elvm/home?authuser=0).
- [2024/03/19] Online demo of EfficientViT-SAM is available: [https://evitsam.hanlab.ai/](https://evitsam.hanlab.ai/). 
- [2024/02/07] We released [EfficientViT-SAM](https://arxiv.org/abs/2402.05008), the first accelerated SAM model that matches/outperforms SAM-ViT-H's zero-shot performance, delivering the SOTA performance-efficiency trade-off.
- [2023/11/20] EfficientViT is available in the [NVIDIA Jetson Generative AI Lab](https://www.jetson-ai-lab.com/tutorial_efficientvit.html).
- [2023/09/12] EfficientViT is highlighted by [MIT home page](https://www.mit.edu/archive/spotlight/efficient-computer-vision/) and [MIT News](https://news.mit.edu/2023/ai-model-high-resolution-computer-vision-0912).
- [2023/07/18] EfficientViT is accepted by ICCV 2023.

## Content

### [ICLR 2025] Deep Compression Autoencoder for Efficient High-Resolution Diffusion Models [[paper](https://arxiv.org/abs/2410.10733)] [[readme](applications/dc_ae/README.md)] [[poster](assets/dc_ae_poster.pdf)]

**Deep Compression Autoencoder (DC-AE) is a new family of high-spatial compression autoencoders with a spatial compression ratio of up to 128 while maintaining reconstruction quality. It accelerates all latent diffusion models regardless of the diffusion model architecture.**

#### Demo

![demo](https://huggingface.co/mit-han-lab/dc-ae-f64c128-in-1.0/resolve/main/assets/dc_ae_demo.gif)
<p align="center">
<b> Figure 1: We address the reconstruction accuracy drop of high spatial-compression autoencoders.
</p>

![demo](assets/dc_ae_diffusion_demo.gif)
<p align="center">
<b> Figure 2: DC-AE speeds up latent diffusion models.
</p>

<p align="center">
<img src="https://huggingface.co/mit-han-lab/dc-ae-f64c128-in-1.0/resolve/main/assets/dc_ae_sana.jpg"  width="1200">
</p>

<p align="center">
<b> Figure 3: DC-AE enables efficient text-to-image generation on the laptop: <a href="https://nvlabs.github.io/Sana/">SANA</a>.
</p>

- [Usage of Deep Compression Autoencoder](applications/dc_ae/README.md#deep-compression-autoencoder)
- [Usage of DC-AE-Diffusion](applications/dc_ae/README.md#efficient-diffusion-models-with-dc-ae)
- [Evaluate Deep Compression Autoencoder](applications/dc_ae/README.md#evaluate-deep-compression-autoencoder)
- [Demo DC-AE-Diffusion Models](applications/dc_ae/README.md#demo-dc-ae-diffusion-models)
- [Evaluate DC-AE-Diffusion Models](applications/dc_ae/README.md#evaluate-dc-ae-diffusion-models)
- [Train DC-AE-Diffusion Models](applications/dc_ae/README.md#train-dc-ae-diffusion-models)
- [Reference](applications/dc_ae/README.md#reference)

### [CVPR 2024 eLVM Workshop] EfficientViT-SAM: Accelerated Segment Anything Model Without Accuracy Loss [[paper](https://arxiv.org/abs/2402.05008)] [[online demo](https://evitsam.hanlab.ai/)] [[readme](applications/efficientvit_sam/README.md)]

**EfficientViT-SAM is a new family of accelerated segment anything models by replacing SAM's heavy image encoder with EfficientViT. It delivers a 48.9x measured TensorRT speedup on A100 GPU over SAM-ViT-H without sacrificing accuracy.**

<p align="left">
<img src="https://huggingface.co/mit-han-lab/efficientvit-sam/resolve/main/sam_zero_shot_coco_mAP.png"  width="500">
</p>

- [Pretrained EfficientViT-SAM Models](applications/efficientvit_sam/README.md#pretrained-efficientvit-sam-models)
- [Usage of EfficientViT-SAM](applications/efficientvit_sam/README.md#usage)
- [Evaluate EfficientViT-SAM](applications/efficientvit_sam/README.md#evaluation)
- [Visualize EfficientViT-SAM](applications/efficientvit_sam/README.md#visualization)
- [Deploy EfficientViT-SAM](applications/efficientvit_sam/README.md#deployment)
- [Train EfficientViT-SAM](applications/efficientvit_sam/README.md#training)
- [Reference](applications/efficientvit_sam/README.md#reference)

### [ICCV 2023] EfficientViT-Classification [[paper](https://arxiv.org/abs/2205.14756)] [[readme](applications/efficientvit_cls/README.md)]

**Efficient image classification models with EfficientViT backbones.**

<p align="left">
<img src="https://huggingface.co/han-cai/efficientvit-cls/resolve/main/efficientvit_cls_results.png"  width="600">
</p>

- [Pretrained EfficientViT Classification Models](applications/efficientvit_cls/README.md#pretrained-efficientvit-classification-models)
- [Usage of EfficientViT Classification Models](applications/efficientvit_cls/README.md#usage)
- [Evaluate EfficientViT Classification Models](applications/efficientvit_cls/README.md#evaluation)
- [Export EfficientViT Classification Models](applications/efficientvit_cls/README.md#export)
- [Train EfficientViT Classification Models](applications/efficientvit_cls/README.md#training)
- [Reference](applications/efficientvit_cls/README.md#reference)

### [ICCV 2023] EfficientViT-Segmentation [[paper](https://arxiv.org/abs/2205.14756)] [[readme](applications/efficientvit_seg/README.md)]

**Efficient semantic segmantation models with EfficientViT backbones.**

![demo](assets/cityscapes_l1.gif)

- [Pretrained EfficientViT Segmentation Models](applications/efficientvit_seg/README.md#pretrained-efficientvit-segmentation-models)
- [Usage of EfficientViT Segmentation Models](applications/efficientvit_seg/README.md#usage)
- [Evaluate EfficientViT Segmentation Models](applications/efficientvit_seg/README.md#evaluation)
- [Visualize EfficientViT Segmentation Models](applications/efficientvit_seg/README.md#visualization)
- [Export EfficientViT Segmentation Models](applications/efficientvit_seg/README.md#export)
- [Reference](applications/efficientvit_seg/README.md#reference)

### EfficientViT-GazeSAM [[readme](applications/efficientvit_gazesam/README.md)]

**Gaze-prompted image segmentation models capable of running in real time with TensorRT on an NVIDIA RTX 4070.**

![GazeSAM demo](https://huggingface.co/mit-han-lab/efficientvit-sam/resolve/main/gazesam/efficientvit_gazesam_demo.gif)

## Getting Started

```bash
conda create -n efficientvit python=3.10
conda activate efficientvit
pip install -U -r requirements.txt
```

## Third-Party Implementation/Integration

- [NVIDIA Jetson Generative AI Lab](https://www.jetson-ai-lab.com/tutorial_efficientvit.html)
- [timm](https://github.com/huggingface/pytorch-image-models): [link](https://github.com/huggingface/pytorch-image-models/blob/main/timm/models/efficientvit_mit.py)
- [X-AnyLabeling](https://github.com/CVHub520/X-AnyLabeling): [link](https://github.com/CVHub520/X-AnyLabeling/blob/main/anylabeling/services/auto_labeling/efficientvit_sam.py)
- [Grounding DINO 1.5 Edge](https://github.com/IDEA-Research/Grounding-DINO-1.5-API): [link](https://arxiv.org/pdf/2405.10300)

## Contact

[Han Cai](http://hancai.ai/)

## Reference

If EfficientViT or EfficientViT-SAM or DC-AE is useful or relevant to your research, please kindly recognize our contributions by citing our paper:

```bibtex
@inproceedings{cai2023efficientvit,
  title={Efficientvit: Lightweight multi-scale attention for high-resolution dense prediction},
  author={Cai, Han and Li, Junyan and Hu, Muyan and Gan, Chuang and Han, Song},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={17302--17313},
  year={2023}
}
```

```bibtex
@article{zhang2024efficientvit,
  title={EfficientViT-SAM: Accelerated Segment Anything Model Without Performance Loss},
  author={Zhang, Zhuoyang and Cai, Han and Han, Song},
  journal={arXiv preprint arXiv:2402.05008},
  year={2024}
}
```

```bibtex
@article{chen2024deep,
  title={Deep Compression Autoencoder for Efficient High-Resolution Diffusion Models},
  author={Chen, Junyu and Cai, Han and Chen, Junsong and Xie, Enze and Yang, Shang and Tang, Haotian and Li, Muyang and Lu, Yao and Han, Song},
  journal={arXiv preprint arXiv:2410.10733},
  year={2024}
}
```