DreamShaper 8 LCM — LocalMuse Core ML
A reproducible Core ML conversion of
Lykon/dreamshaper-8-lcm for on-device image generation in the LocalMuse iOS app.
This is a format conversion, not a retrained model. Lykon remains the creator of DreamShaper 8 LCM. The UNet uses Apple 6-bit k-means palettization; the matching CLIP text encoder and both VAE directions remain FP16 to preserve prompt and image fidelity.
Pinned source
- Repository:
Lykon/dreamshaper-8-lcm
- Revision:
4645d8bc6a8e6b106d21606d63e8460cdad4f1a6
- Checkpoint:
DreamShaper8_LCM.safetensors
- Size:
2,133,804,992 bytes
- SHA-256:
a4f3e1526c5dc4fcbe342f5c410d83ae202c7a415fcefcbb92e0f93fcd0a87c3
The SD 1.5 component configuration and tokenizer are pinned to stable-diffusion-v1-5/stable-diffusion-v1-5@451f4fe16113bff5a5d2269ed5ad43b0592e9a14. No additional LCM LoRA is fused because the source checkpoint is already LCM-distilled.
Conversion
- Apple
ml-stable-diffusion revision: e12202c1f6405b83918b58a5d097cd61e3e1f702
- Attention:
SPLIT_EINSUM_V2
- UNet: 6-bit palettized, monolithic
[2, 4, 64, 64] -> [2, 4, 64, 64]
- Text encoder: FP16
[1, 77] -> [1, 77, 768]
- VAE encoder and decoder: FP16, native 512×512
- Minimum compiled deployment target: iOS 17
A deterministic source-to-compiled comparison produced:
- UNet: correlation
0.99994064, relative RMSE 0.011402
- Text encoder: correlation
0.99999909, relative RMSE 0.001342
PROVENANCE.json authenticates every source and generated artifact. Every downloadable file is also independently pinned by size and SHA-256 in the LocalMuse catalog.
Recommended settings
DreamShaper 8 LCM must run with an LCM scheduler:
- Steps: 5–15 (default 10)
- CFG scale: approximately 2
- Resolution: 512×512
As noted by the original author, LCM is optimized for speed and has lower maximum quality than the full DreamShaper 8 checkpoint.
License and attribution
The converted weights remain subject to the included CreativeML Open RAIL-M license and all use-based restrictions. See the original
DreamShaper 8 LCM model card for upstream attribution and model information.
The files were modified from the source by conversion to compiled Apple Core ML format and by 6-bit palettization of the UNet. No additional usage restrictions are imposed by this repository.
