1-Bit and Ternary Bonsai Image 4B: Image Generation for Local Devices (prismml.com)

The introduction of Bonsai Image 4B marks a significant advancement in local image generation, offering two compact models—1-bit and Ternary—specifically designed to deliver high-quality diffusion inference on devices such as smartphones and laptops. The 1-bit variant uses binary transformer weights and achieves a compression ratio that reduces its footprint to just 0.93 GB, while the Ternary variant employs both binary and zero states to enhance visual fidelity, reaching a size of 1.21 GB. This makes it possible to run powerful models like these directly on devices like the iPhone 17 Pro Max, which previously could not accommodate such complex architectures. This development is pivotal for the AI/ML community as it shifts local image generation towards practicality and accessibility, minimizing reliance on cloud-based services that often bring latency and cost issues. Users can now iterate on images quickly and privately, with 1-bit Bonsai Image retaining 88% and Ternary Bonsai Image maintaining 95% of the accuracy of larger models while drastically reducing memory usage. The models promise to revolutionize how creators engage with image generation, making the workflow more seamless and integrated into user experiences, thus empowering a new era of on-device AI applications. Both variants will be released with open weights and code under the Apache 2.0 license, promoting further exploration and development in the field.