LocalMuseAI/coreml-dreamshaper-xl-lightning-6bit-v2

DreamShaper XL Lightning — LocalMuse Core ML

Reproducible Core ML build for on-device inference in LocalMuse. Generation remains entirely local after download.

• Source: Lykon/dreamshaper-xl-lightning at pinned revision 561086c22aba3925b1e8803a73857f061635c7c3.
• Native resolution: 1024×1024.
• Workflows: Text to Image and Image to Image.
• Recommended inference: DPM-Solver++ Multistep, leading timestep spacing, 4 steps, CFG 2.0.
• UNet: model-specific 6-bit palettization, split into two sub-1 GiB Core ML programs, SPLIT_EINSUM attention.
• Text encoders: both taken from the same pinned DreamShaper snapshot; CLIP-L stays FP16 and OpenCLIP-bigG uses 8-bit palettization. No encoder is reused from another model.
• VAE: matching source VAE encoder/decoder at FP16, scaling factor 0.13025.
• Device class: iPhone with nominal 8 GB memory or newer supported class.

PROVENANCE.json records every source file hash, conversion dependency, tensor contract, validation result and output hash. The conversion passed PyTorch/Core ML component parity checks and an end-to-end compiled Core ML 1024×1024 generation test. Final Neural Engine validation is performed on iPhone because this model's split UNet is not executable through the macOS ANE backend used by the conversion host.

This repository contains converted model weights. The upstream CreativeML Open RAIL++-M terms and use restrictions continue to apply. See LICENSE.md and the source model card.