LocalMuseAI/coreml-realvisxl-v5-lightning-6bit-v2

RealVisXL V5.0 Lightning — exact-source 6-bit split Core ML

Self-contained SDXL package prepared for on-device use by the LocalMuse iOS app. Unlike the previous LocalMuse artifact, every learned component in this revision comes from the same pinned RealVisXL snapshot: UNet, both text encoders, VAE encoder and VAE decoder. The tokenizer files are also copied from that snapshot.

Authenticated source

• Repository: SG161222/RealVisXL_V5.0_Lightning
• Revision: f4454158cedaab9f0688c199561d6c92525f3a85
• Snapshot manifest SHA-256: b82b09421cb3cf593f0f57e3a3fa5858ec75409f2830e779f0184efb735d33ee
• Civitai/Hugging Face fp16 checkpoint SHA-256: fabcadd9330dcc4f9702063428d40b9d4d07168d8acefc819b8d1d9db466b3ec
• UNet fp16 SHA-256: 1143cd2aaf65d24af34b5699d090aed724f6c0978c2ec5a5f56821ccb36260ce
• TextEncoder fp16 SHA-256: 0a873801e35a008b17e228189a83a8b41e068ffe7bde7c1105e8c6e947758bc5
• TextEncoder2 fp16 SHA-256: dfa429b268cd3f6297927f1fd3eacac04781b17ab4caa97892a461ff41400a5e
• VAE fp16 SHA-256: 6353737672c94b96174cb590f711eac6edf2fcce5b6e91aa9d73c5adc589ee48
• License: CreativeML Open RAIL++-M (LICENSE.md)

The source checkpoint hash is identical to Civitai RealVisXL V5.0 Lightning (BakedVAE), model version 798204. No checkpoint merge or component substitution is performed by this conversion.

Conversion

• Apple converter revision: e12202c1f6405b83918b58a5d097cd61e3e1f702
• PyTorch 2.4.0, coremltools 8.0, Diffusers 0.30.2
• Fixed resolution: 1024 × 1024 (128 × 128 latent)
• Classifier-free-guidance UNet batch: 2
• Attention: SPLIT_EINSUM
• UNet: 6-bit palettized, then split into two chunks below 1 GiB
• CLIP-L TextEncoder: FP16
• OpenCLIP-bigG TextEncoder2: 8-bit palettized
• VAE encoder and decoder: converted from the pinned RealVisXL VAE
• Minimum deployment target: iOS 17

Author-aligned runtime profile

• Sampler: DPM++ SDE Karras (single-step SDE with midpoint evaluation)
• Steps: 5 by default; 4–6 recommended
• CFG: 2.0 by default; 1.0–2.0 recommended
• Native generation: 1024 × 1024

Five requested scheduler steps execute nine UNet evaluations. This is not the ordinary DPM-Solver++ multistep scheduler and not DPM++ 2M SDE.

The author's reference images may include a separate Hi-Res pass. That pass is not baked into these model files and must be implemented by the client when enabled.

Validation

• VAE decoder parity: 106.3 dB
• VAE encoder parity: 101.5 dB
• TextEncoder parity: 81.1 dB
• TextEncoder2 parity before 8-bit storage conversion: 55.2 dB
• UNet Core ML parity before 6-bit storage conversion: 45.7 dB
• Compiled contracts: UNet [2, 4, 128, 128] → [2, 4, 128, 128], VAE 1024 × 1024, both text encoders 77 tokens
• Every compiled component loads successfully through Core ML on macOS with CPU-only compute. Final Neural Engine quality, memory and thermal validation must be performed on a supported 8 GiB iPhone before release.