bytevisionlab/dreamlite

DreamLite is a research project that packs image generation and image editing into one small AI model, small enough to run directly on a phone with no internet connection needed. You type a text description, such as a close-up of a dragon breathing fire, or an instruction for editing an existing photo, such as changing the background to a forest, and the model produces a 1024 by 1024 pixel image in about 3 seconds on an iPhone 17 Pro. The model has 0.39 billion parameters, which is small compared to most image generation models, and it comes in two versions. The base version takes 28 steps to produce a higher quality image, while the mobile version is tuned for speed and only needs 4 steps. Both versions handle text-to-image generation and text-guided editing using the same underlying architecture, rather than needing two separate models for the two tasks. You can try the model through a command line interface, passing in a text prompt and, for editing tasks, a path to a source image, or through a Gradio-based web interface that runs locally or is hosted on Hugging Face Spaces. The repository also includes benchmark scripts for comparing DreamLite's results against other image generation and editing models on standard test sets. The model weights themselves are not open for immediate download. They are undergoing a safety review, and you need to email the authors with your name, affiliation, and intended use case to request early access. The README also states usage guidelines: the model must not be used to create sexually explicit, violent, discriminatory, or otherwise illegal content. The project is backed by a published research paper and comes from ByteVisionLab.