zimageapp/MOSS-Image-SFT-MLX-Q4

MOSS-Image-SFT MLX Q4 + LoRAs

A complete MOSS-Image-SFT package for Apple Silicon. The base is the 7.47 GB MFlux Q4 conversion. The 4-step Turbo adapter and anime-style adapter are included so the app can verify and download all required files from one repository.

Use in Z Image

Choose MOSS IMAGE in the sidebar and download from Hugging Face or ModelScope. The default destination is /Volumes/Z1/huggingface/hub/models/MOSS-Image-SFT-MLX-Q4. The app uses the pinned MFlux MLX implementation and applies selected adapters before sampling. Both adapters are enabled by default. 4-Step Turbo runs exactly 4 steps at guidance 1; turn it off to use the standard 50-step sampler. Anime style uses scale 1.

Use widths and heights divisible by 16. The prompt accepts natural language, tags, and JSON layouts. A 256×256 local smoke generation with the base Q4 model used about 7.6 GB peak MLX memory; actual working memory varies with image size and other apps.

Contents

• transformer/, text_encoder/, tokenizer/, vae/: MFlux Q4 conversion of OpenMOSS-Team/MOSS-Image-SFT, from mflux-community/moss-image-sft-mflux-q4.
• loras/moss-image-sft-4step.safetensors: akhaliq/MOSS-Image-SFT-4-steps.
• loras/moss-image-anime.safetensors: axiomofmind/Moss-image-anime-lora.
• package-manifest.json: sizes and SHA-256 hashes used by the app downloader.

All bundled weights are Apache-2.0. The combined package is unofficial. See LICENSE, NOTICE, licenses/, and source-manifest.json.