litert-community/Bonsai-Image-ternary-4B

Card: source link to prism-ml/bonsai-image-ternary-4B-unpacked, jinja2 in the pip line, CompiledModel API notes in place of the Interpreter snippet

generate.py: CompiledModel API (same as litert-samples)

Card: definition line (LiteRT, 2026-09)

Card: base_model_relation: quantized (list under the base model's Quantizations)

Add measured Galaxy S26 NPU/GPU section

Card: collect the measured timings into a Performance table with their conditions

Card: add the GPU-shaped DiT (dit_gpu_int4b32) + macOS Metal section

Add GPU-shaped DiT export (dit_gpu_int4b32) for the macOS Metal demo app

Model card: on-device run section (Android/iOS apps + minimal Kotlin)

Model card: link the litert-samples sample (models/bonsai/bonsai_image_4b)

Bonsai Image 4B LiteRT: int4 DiT + int4/int8wo text encoder + fp32 VAE, torch-free sample, iPhone-verified

initial commit