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

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

Self-contained, compiled SDXL package prepared for on-device use by the LocalMuse iOS app. The repository stays private until the artifact manifest and anonymous downloads are verified.

Provenance

• Source UNet: SG161222/RealVisXL_V5.0_Lightning
• Pinned source revision: f4454158cedaab9f0688c199561d6c92525f3a85
• Source fp16 UNet LFS SHA-256: 1143cd2aaf65d24af34b5699d090aed724f6c0978c2ec5a5f56821ccb36260ce
• License: CreativeML Open RAIL++-M (LICENSE.md), inherited from the source model and SDXL base.

The RealVisXL UNet was converted from the pinned source above. The text encoders, VAE encoder/decoder, tokenizer vocabulary, and merges are standard SDXL components reused byte-for-byte from LocalMuseAI/coreml-dreamshaper-xl-lightning-6bit revision 736408cecf59b3b81a8b2a900cdd5dbf94c34c2a. Only the UNet contains RealVisXL-specific weights.

Conversion

• Apple converter: apple/ml-stable-diffusion revision e12202c1f6405b83918b58a5d097cd61e3e1f702
• PyTorch 2.4.0
• coremltools 8.0
• Diffusers 0.30.2
• Transformers 4.44.2
• huggingface_hub 0.24.6
• NumPy 1.23.5
• Resolution: fixed 1024 × 1024 (128 × 128 latent)
• Classifier-free-guidance batch: 2
• Attention implementation: SPLIT_EINSUM
• UNet storage: mixed Float16 and 6-bit palettized weights
• Minimum Core ML deployment target: iOS 17
• UNet split into two independently compiled chunks below 1 GiB each

The source model's fp16 UNet was loaded directly without downloading unrelated full-precision weights, converted through Apple's Stable Diffusion implementation, palettized before chunking, bisected with coremltools, and compiled with xcrun coremlc.

Runtime settings

The source model card recommends 5 inference steps, CFG 1–2, and a DPM++ SDE Karras scheduler. LocalMuse uses:

• Steps: 5
• CFG: 1.5
• Scheduler: second-order DPM-Solver++ multistep
• Timestep spacing: Karras
• Batch size: 1 image (the UNet internally uses batch 2 for CFG)
• UNet: Neural Engine preferred
• VAE: GPU preferred
• Minimum device memory exposed by the app: 8 GiB

Verification

Both chunks compile successfully with Xcode coremlc. Their named input/output feature sets and tensor shapes match the SDXL split-UNet contract already used by LocalMuse. The model is distributed as direct authenticated files; the app pins every URL to a repository commit and verifies the exact SHA-256 and byte count before installation.

Real-device image quality and thermal behavior should still be validated before shipping a production App Store build.