We introduce DreamLite, a compact and unified on-device diffusion model (0.39B) that seamlessly supports both text-to-image generation and text-guided image editing within a single network architecture.

Built upon a pruned mobile U-Net backbone, DreamLite unifies multimodal conditioning through In-Context Spatial Concatenation directly in the latent space. By leveraging progressive step distillation, DreamLite achieves ultra-fast 4-step inference, capable of generating or editing a 1024×1024 image in ~3 seconds on an iPhone 17 Pro (powered by 4-bit Qwen-VL and fp16 VAE+UNet) — operating fully on-device with zero cloud dependency.

Figure 1. The overall unified architecture of DreamLite.

• [2026.06] 🎉🎉🎉 DreamLite has been merged into the official 🤗 Diffusers library (pending the next release). You can now load and run DreamLite directly via diffusers.
• [2026.06] 🎉🎉🎉 Community contribution: @ENUMERA8OR released dreamlite-comfyui-lowvram — ComfyUI custom nodes and a low-VRAM inference pipeline that runs DreamLite-base at 1024×1024 on 4 GB GPUs.
• [2026.04] 🎉🎉🎉 We officially released the inference code.

• [2026.03] 🎉🎉🎉 DreamLite is publicly announced! Check out our project page and arXiv paper.

Experience real-time generation and editing on an iPhone 17 Pro. No internet connection or cloud processing required.

Note: If demos fail to render natively on GitHub, please visit our Project Page to watch the full demonstrations.

# Clone the repository
git clone https://github.com/ByteVisionLab/DreamLite.git
cd DreamLite

# Create and activate a conda environment
conda create -n dreamlite python=3.10 -y
conda activate dreamlite

# Install dependencies
pip install -r requirements.txt

Download the weights from DreamLite-base (28 steps) or DreamLite-mobile (4 steps). Both model repositories have two branches, each corresponding to a different inference implementation:

The branches use different model configurations and module paths. Choose the branch that matches your inference code, and keep downloads from the two branches in separate directories.

For this repository's CLI scripts, download the main branch:

hf download carlofkl/DreamLite-base --revision main --local-dir models/DreamLite-base
hf download carlofkl/DreamLite-mobile --revision main --local-dir models/DreamLite-mobile

The scripts use the following local paths by default:

DreamLite/
├── models/
│   ├── DreamLite-base/
│   └── DreamLite-mobile/

DreamLite has been merged into the official 🤗 Diffusers library. Since the change is not yet included in a stable Diffusers release, please install the latest main branch directly from source:

pip install git+https://github.com/huggingface/diffusers.git

Use the diffusers weight branch listed above. from_pretrained(..., revision="diffusers") automatically downloads this branch on first use. For Base, import DreamLitePipeline; for Mobile, import DreamLiteMobilePipeline.

Alternatively, pre-download the Diffusers weights into separate local directories:

hf download carlofkl/DreamLite-base --revision diffusers --local-dir models/DreamLite-base-diffusers
hf download carlofkl/DreamLite-mobile --revision diffusers --local-dir models/DreamLite-mobile-diffusers

Then load from the corresponding local path, for example: DreamLitePipeline.from_pretrained("models/DreamLite-base-diffusers", torch_dtype=dtype).

Then you can load and run DreamLite with just a few lines of code:

import torch
from diffusers import DreamLitePipeline
from diffusers.utils import load_image

model_id = "carlofkl/DreamLite-base"
device = "cuda"
dtype = torch.float16

pipe = DreamLitePipeline.from_pretrained(model_id, revision="diffusers", torch_dtype=dtype)
pipe.to(device=device)

# Text-to-image
image = pipe(
    prompt="A serene mountain lake at sunrise",
    generator=torch.Generator(device=device).manual_seed(42),
).images[0]

image.save("dreamlite_t2i.png")

# Image-to-image (instruction-based edit)
image_url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg"
init_image = load_image(image_url)
edited = pipe(
    prompt="make it snowy",
    image=init_image,
    generator=torch.Generator(device=device).manual_seed(42),
).images[0]

edited.save("dreamlite_i2i.png")

These scripts import the local dreamlite package and use the main weight branch downloaded in Section 1.

# ==========================================
# DreamLite-base: 28 Steps (High Fidelity)
# ==========================================
# Text-to-Image Generation
python infer.py --prompt "A close-up of a fire spitting dragon cinematic shot."

# Text-guided Image Editing
python infer.py --prompt "Transfer this image to oil-painting style." --image_path ./inputs/source.png

# ==========================================
# DreamLite-mobile: 4 Steps (Ultra Fast)
# ==========================================
# Text-to-Image Generation
python infer_mobile.py --prompt "A portrait of a young woman with flowers." 

# Text-guided Image Editing
python infer_mobile.py --prompt "Change the background to a dense forest." --image_path ./inputs/source.png

We provide comprehensive benchmark evaluation scripts (GenEval & ImgEdit) to facilitate performance comparisons between DreamLite and other state-of-the-art models. Please configure your local dataset paths within tools/benchmark/infer_geneval.py and tools/benchmark/infer_imgedit.py prior to execution.

# Run the benchmark evaluation
python tools/benchmark/infer_geneval.py --save_dir ./output/benchmark/geneval_output --geneval_json "YOUR_GENEVAL/evaluation_metadata.jsonl"
python tools/benchmark/infer_imgedit.py --save_dir ./output/benchmark/imgedit_output --json_path "YOUR_IMGEDIT_PATH/ImgEdit/Benchmark/Basic/basic_edit.json" --img_root "YOUR_IMGEDIT_IMAGES_PATH/ImgEdit/Benchmark/singleturn"

We provide a user-friendly web interface powered by Gradio. You can try our live demo on Hugging Face Spaces, or deploy it locally on your own machine (GPU/CPU).

To run the interactive demo locally:

# Launch the local web server
python tools/app.py

We offer two distinct variants of the DreamLite model to provide an optimal balance between visual fidelity and on-device inference latency.

Weights are available on Hugging Face: DreamLite-base and DreamLite-mobile. See Model weights and branches to choose the branch for your inference code.

⚠️ Important Usage and Compliance Notice: By accessing and using these models, you agree to abide by our ethical guidelines. These models are for non-commercial, research-only use. You must NOT use them to generate, edit, or distribute any content that is sexually explicit, pornographic, violent, discriminatory, or otherwise illegal. Commercial use and public redistribution of the model weights are strictly prohibited.

Quantitative comparison with state-of-the-art methods on generation and editing benchmarks.

Text-to-Image generation comparison.

Text-guided image editing comparison.

We provide comprehensive support for LoRA fine-tuning and inference, enabling lightweight customization of DreamLite on your own domain-specific datasets.

For detailed instructions, training scripts, and examples, please refer to our dedicated LoRA Fine-Tuning Guide.

We provide a complete iOS On-device Deployment Reference, including model export scripts (CoreML + mlx-vlm 4-bit quantization), modified Swift library files, and iOS app source code.

We sincerely thank the community for extending DreamLite to broader use cases and hardware. If you have built something on top of DreamLite, feel free to open a PR/Issue and we'd be happy to feature it here.

• dreamlite-comfyui-lowvram by @ENUMERA8OR — ComfyUI custom nodes and a low-VRAM inference pipeline that runs DreamLite-base at 1024×1024 on a 4 GB NVIDIA GPU (e.g., GTX 1650 Ti). It enables this through sequential CPU offload, float32 precision, and a GQA-aware query-token attention slicing strategy that preserves DreamLite's grouped-query attention layout (which standard Diffusers attention slicing would break). The repository also bundles ComfyUI workflows for both DreamLite-base and DreamLite-mobile, plus tiled RealESRGAN upscaling support.

• Release paper on arXiv
• Release inference code
• Release LoRA training
• Release model weights on HuggingFace
• Release online demo
• On-device Deployment Reference

We thank the great work from SDXL, SnapGen, Qwen and TAESDXL. The work is under supervision from Prof. Wangmeng Zuo.

Code: Apache-2.0

Model weights: see WEIGHTS_LICENSE, CC BY-NC 4.0

If our work assists your research, feel free to give us a star ⭐ or cite us using:

@article{feng2026dreamlite,
  title={DreamLite: A Lightweight On-Device Unified Model for Image Generation and Editing},
  author={Kailai Feng and Yuxiang Wei and Bo Chen and Yang Pan and Hu Ye and Songwei Liu and Chenqian Yan and Yuan Gao},
  journal={arXiv preprint arXiv:2603.28713},
  year={2026}
}