Dadm-n/stable-diffusion-v1-5-lcm-inpainting-coreml

Stable Diffusion 1.5 LCM Inpainting — CoreML (SHxTLxST)

CoreML .mlpackage versions of SimianLuo/LCM_Dreamshaper_v7 configured for inpainting, palettized per Apple's LCM recommendation (TextEncoder/VAE at 6-bit, UNet at 8-bit).

Used by SHxTLxST iOS app for lifestyle image generation from masked product photos.

Submodels

Submodel	nbits	Size	Purpose
TextEncoder.mlpackage	6-bit	88 MB	CLIP text encoding
Unet.mlpackage	8-bit	821 MB	Latent diffusion (inpainting)
VAEEncoder.mlpackage	6-bit	25 MB	Image -> latent
VAEDecoder.mlpackage	6-bit	36 MB	Latent -> image

Total: 970 MB (well within iOS app size budgets)

Usage (iOS, Swift)

import CoreML

let config = MLModelConfiguration()
config.computeUnits = .all  // ANE + GPU + CPU

let textEncoder = try TextEncoder(configuration: config)
let unet = try Unet(configuration: config)
let vaeEncoder = try VAEEncoder(configuration: config)
let vaeDecoder = try VAEDecoder(configuration: config)

For inference, see apple/ml-stable-diffusion — this package is drop-in compatible, with the UNet modified to accept a 9-channel input (4 latent + 4 masked image latent + 1 mask) per the inpainting convention.

Files

• Resources/ — the four .mlpackage submodels
• tokenizer/ — CLIP tokenizer (vocab.json + merges.txt)
• scheduler_config.json — LCM scheduler config
• manifest.json — per-submodel quantization metadata

Source

• Original PyTorch model: SimianLuo/LCM_Dreamshaper_v7
• Base: runwayml/stable-diffusion-v1-5
• Conversion script: convert_sd_inpaint_apple.py
• Conversion pipeline: trace on Linux+GPU -> convert on macOS (coremltools BlobWriter is macOS-only)

License

Original SD 1.5: CreativeML Open RAIL-M. The CoreML conversion is a derivative work; this repo inherits the Open RAIL-M license.