coreml-cyberrealistic-xl-v10-6bit

CyberRealistic XL V10 — Core ML for LocalMuse

A production-oriented Core ML conversion of Cyberdelia's native, non-distilled CyberRealistic XL Play V10.0 FP16 checkpoint for on-device 1024×1024 Text to Image generation in LocalMuse.

Runtime profile

• Native SDXL 1024×1024; this is not Lightning or LCM.
• 25–40 diffusion steps; LocalMuse defaults to 30.
• DPM-Solver++ Karras, CFG 7.5, batch size one.
• Two 6-bit split UNet components, each below the iOS 1 GiB component limit.
• CyberRealistic V10 CLIP-L at FP16 and OpenCLIP-bigG at 8-bit.
• Checkpoint's built-in VAE decoder and encoder; the encoder is included for LocalMuse Face Detail.
• Requires a device in the nominal 8 GiB memory class.

The 6-bit UNet and 8-bit second text encoder are storage palettization. They do not turn the checkpoint into a reduced-step model or change its normal diffusion schedule.

Source and reproducibility

Source checkpoint: cyberdelia/CyberRealisticXL, revision c04ae13b465d19763561a834d625a3d504a4f315, file CyberRealisticXLPlay_V10.0_FP16.safetensors, SHA-256 fd5e870b5bbce4bddeb64f4bb8e49c57f84ab793c0262a503f0123be435e667d.

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 SDXL Base FP16 layer. SDXL prompt conditioning uses the preceding hidden state, while pooled conditioning is produced by the second encoder. Exact inputs, revisions, hashes, conversion versions, component sizes, and validation results are recorded in PROVENANCE.json.

Validation

Every compiled component was executed with finite outputs. The complete split-UNet Swift pipeline then generated a 1024×1024 image with 25 steps, DPM-Solver++ Karras, CFG 7.5, and a fixed seed.

License and attribution

CyberRealistic XL is by Cyberdelia and is published with creativeml-openrail-m metadata. The full license text is included in LICENSE.md. The license and any use-based restrictions continue to apply to this converted artifact.