---
license: cc-by-nc-4.0
tags:
  - dreamlite
  - dreamlite-mobile
  - onnx
  - webgpu
  - text-to-image
  - diffusion
  - fp16
  - browser
library_name: onnx
pipeline_tag: text-to-image
base_model: ByteVisionLab/DreamLite
---

# DreamLite-Mobile ONNX (WebGPU browser bundle)

Browser-ready **ONNX fp16** weights for running **DreamLite-Mobile** diffusion in the web via **ONNX Runtime Web** and **WebGPU**. This repo is the **Option A** bundle: UNet + VAE only (~**750 MB** uncompressed).

## What is in this repository

| File | Approx. size | Role |
|------|-------------:|------|
| `unet_fp16.onnx` | ~746 MB | 4-step distilled diffusion UNet |
| `vae_decoder_fp16.onnx` | ~2.4 MB | VAE decode (latents to image) |
| `vae_encoder_fp16.onnx` | ~2.4 MB | VAE encode (image to latents) |

- **4-step** DreamLite-Mobile distilled sampling (match your app scheduler settings to the mobile export).
- **Text encoder is not included.** You still need:
  - **`encode_server`** (or equivalent) to produce text embeddings at runtime, and
  - The full **DreamLite-Mobile** setup / weights — see [carlofkl/DreamLite-mobile](https://huggingface.co/carlofkl/DreamLite-mobile).

## Base model

Derived from **[ByteVisionLab/DreamLite](https://huggingface.co/ByteVisionLab/DreamLite)** (DreamLite family). This upload packages the mobile ONNX slice for in-browser inference, not the full training checkpoint.

## Using these files in a web UI

1. **Local / static hosting:** copy the three `.onnx` files into your app `public/models/` (or another static URL your bundler serves).
2. **Configurable base URL:** set **`VITE_MODEL_BASE`** to the directory URL where the three files live (trailing slash behavior depends on your app; point at the folder that contains the `.onnx` names above).

Example (conceptual):

```env
VITE_MODEL_BASE=https://huggingface.co/dror201031/DreamLite-mobile-ONNX-WebGPU/resolve/main/
```

Your frontend should request `unet_fp16.onnx`, `vae_decoder_fp16.onnx`, and `vae_encoder_fp16.onnx` under that base.

## What is not included

- Text encoder ONNX
- Validation images or test fixtures
- `node_modules` or application source

## License

**CC BY-NC 4.0** — consistent with the DreamLite model family. See the [license](https://creativecommons.org/licenses/by-nc/4.0/) for commercial use restrictions.

## Citation / links

- Base: [ByteVisionLab/DreamLite](https://huggingface.co/ByteVisionLab/DreamLite)
- Mobile weights reference: [carlofkl/DreamLite-mobile](https://huggingface.co/carlofkl/DreamLite-mobile)