| --- |
| license: openrail++ |
| base_model: SG161222/RealVisXL_V5.0_Lightning |
| library_name: coreml |
| tags: |
| - coreml |
| - stable-diffusion-xl |
| - text-to-image |
| - image-to-image |
| - ios |
| - localmuse |
| --- |
| |
| # 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. |
|
|
