coreml-wai-illustrious-sdxl-v17-6bit

WAI Illustrious SDXL V17 — Core ML for LocalMuse

A production-oriented Core ML conversion of WAI Illustrious SDXL V17 for private, on-device 1024×1024 generation in LocalMuse.

Runtime profile

• Native, non-distilled Illustrious SDXL at 1024×1024.
• 15–30 diffusion steps; LocalMuse defaults to 25.
• DPM-Solver++ Karras, CFG 5–7, batch size one.
• Recommended positive quality tags: masterpiece, best quality, amazing quality.
• Two 6-bit split UNet components, each below the iOS 1 GiB component limit.
• Model CLIP-L at FP16 and OpenCLIP-bigG at 8-bit.
• Checkpoint VAE decoder and encoder for Text to Image, Transform, and LocalMuse Face Detail.
• Requires an iPhone or iPad in the nominal 8 GiB memory class.

The UNet and second text-encoder palettization reduce model storage. They do not turn V17 into a Lightning, LCM, Turbo, or other reduced-step model.

Source and reproducibility

Official release: WAI Illustrious SDXL on Civitai, model ID 827184, version ID 2883731.

Conversion source checkpoint: LyliaEngine/waiIllustriousSDXL_v170, revision 5ef4e2da7173a160ad04aebcaa2fdcd6d20ed792, file waiIllustriousSDXL_v170.safetensors, SHA-256 f116b0c78ff441467b0cdc8f1936e1ed18ea31e9997c7b132b1b8db533f0bd04.

The checkpoint contains non-finite values only in CLIP-L's unused final transformer layer. The converter replaces only those values with the matching pinned Illustrious base-model layer. SDXL prompt conditioning uses the preceding hidden state, while pooled conditioning is produced by the exact WAI second encoder. Exact inputs, revisions, hashes, conversion versions, component sizes, and validation status are recorded in PROVENANCE.json.

License and attribution

WAI Illustrious SDXL is by WAI0731. The official Civitai release declares the Fair AI Public License 1.0-SD. The complete license text is included in LICENSE.md; its use restrictions continue to apply to this converted artifact. The Hugging Face source is used as a pinned transport mirror and its inconsistent repository metadata does not replace the license declared by the official release.

Validation status

Checkpoint shapes, all tensors used by inference, compiled component shapes, and component size limits were verified. The final iPhone runtime test remains pending and is intentionally not represented as completed in provenance.