Modelpublic

SceneWorks/z-image-mlx

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face

0likes

Model Card

Z-Image — MLX quant-matrix (SceneWorks re-host)

Pre-built MLX (Apple Silicon) quantization tiers of `Tongyi-MAI/Z-Image`, hosted by SceneWorks for direct, ready-to-run loading in the SceneWorks desktop app (no install-time conversion, no gated download).

Tiers

Each subdirectory is a complete, self-contained snapshot (transformer + Qwen3 text encoder + VAE + tokenizer + scheduler) that the SceneWorks z_image engine loads directly:

Tier	Subdir	Precision	Use
Q4 (default)	`q4/`	4-bit group-affine (group 64) weights; dense norms	smallest footprint (undistilled base, real CFG)
Q8	`q8/`	8-bit group-affine weights	higher fidelity
bf16	`bf16/`	dense bf16	maximum fidelity

The transformer, text encoder, and VAE attention are quantized in the Q4/Q8 tiers; the bf16 tier is the full dense model. The packed weights auto-detect their quantization on load (no manifest needed).

License

Apache-2.0, inherited from the upstream Tongyi-MAI/Z-Image. This is an unmodified-weights re-host (re-quantized for MLX). All credit to the Tongyi-MAI team.