coreml-perfectdeliberate-v10-6bit

PerfectDeliberate V10 — LocalMuse Core ML

This repository contains the iOS Core ML conversion used by LocalMuse for PerfectDeliberate V10. It is a fixed 1024×1024 Illustrious/SDXL pipeline with two sub-1 GiB 6-bit UNet chunks, FP16 CLIP-L, 8-bit OpenCLIP-bigG, and the checkpoint's own FP16 VAE decoder and encoder.

The package supports Text to Image, Image to Image, and LocalMuse Face Detail. It is intended for devices in the nominal 8 GiB memory class or newer.

Source and corresponding source offer

• Creator: Desync
• Model card: https://civitai.com/models/24350/perfectdeliberate?modelVersionId=3187704
• Corresponding source checkpoint: https://civitai.com/api/download/models/3187704
• Source filename: perfectdeliberate_v10.safetensors
• Source size: 7105347928 bytes
• Source SHA-256: cbba6875a194e16a0f9e7a812289221fb437d0ebdddf0d99a6876fd0e903ad8e

Full conversion provenance, pinned revisions, tensor contracts, repair scope, and validation results are recorded in PROVENANCE.json.

Recommended LocalMuse settings

• Resolution: 1024×1024
• Sampler: DPM++ 2M Karras
• Steps: 20–50 (default 30)
• CFG: 5–8 (default 7)

The creator also recommends portrait resolutions, but this Core ML artifact is intentionally fixed to the 1024×1024 size supported by LocalMuse's current SDXL runtime.

License

The model and this converted derivative are distributed under the Fair AI Public License 1.0-SD. See LICENSE.md. The license's prohibited-use terms continue to apply.