LocalMuseAI/mlx-z-anime-distill-4step-4bit

Z-Anime Distill 4-Step — MLX int4

Format conversion for LocalMuse of SeeSee21's Z-Anime Distill-4-Step (diffusion_models/z-anime-distill-4step-bf16.safetensors from SeeSee21/Z-Anime at revision 0f5fb51464638a2f7328a1d74590281e63e1fde2). Credit belongs to SeeSee21. This repository only changes the tensor format (int4 MLX); no training, LoRA or model merge was performed here.

License and attribution

Z-Anime is released by its author under Apache 2.0, and it is a fine-tune of Tongyi-MAI/Z-Image (Apache 2.0). The checkpoint's embedded ComfyUI save metadata (preserved under source.* in transformer/model.safetensors) shows the 4-step variant is Z-Anime Base with alibaba-pai/Z-Image-Fun-Lora-Distill Z-Image-Fun-Lora-Distill-4-Steps-2603 (Apache 2.0) merged at strength 1.0 by the author. The Qwen3-4B encoder, tokenizer and 16-channel VAE are the official Tongyi-MAI/Z-Image-Turbo components (Apache 2.0). See LICENSE and source.json.

Modifications: the DiT's attention and feed-forward weights were quantized to affine int4 (group 64) and Q/K/V were split; the encoder was quantized to int4 and truncated to the 35 layers whose penultimate hidden state is consumed.

Settings (author's recommendation)

• Text to Image, batch one, 896 × 1152 (default) or 1024 × 1024.
• Euler / beta (ComfyUI beta_scheduler, alpha = beta = 0.6), 4 steps (up to 6), distilled CFG 1.0, no negative prompt.
• Static flow shift 3.5 (ModelSamplingAuraFlow, from the author's workflow), 1000-point training sigma grid, reversed model time 1-sigma, negative model velocity. runtime_config.json records the shift.
• Natural-language prompts work best.

Tensor contract

Profile zImageTurbo6B4Bit, MLX Swift 0.31.4 (conversion with MLX Python 0.31.2). S3-DiT attention/FFN: affine int4, group 64; embedders, modulation and norms stay BF16. Original 16-channel VAE (scale 0.3611, shift 0.1159) decoded in FP32 overlapping 384-pixel windows. Encoder, transformer and decoder load and evict sequentially.

The eight runtime files total 6,046,965,458 bytes; every size and SHA-256 is in manifest.json. The encoder, tokenizer and VAE bytes are identical to LocalMuse's validated Z-Image profile. Source checkpoint SHA-256: 4a083f389432c5cf04bfa686a1e819f894d01e4a00bdf02e846617e17052ac77.

Validation

Native MLX inference with Metal API validation on an Apple M1 Pro / 16 GiB (Swift Debug build, not iPhone timing):

Canvas	Steps	Seed	Total	Peak MLX allocation
896 × 1152	4	1729	128.35 s	4,760,601,834 B
1024 × 1024	4	42	127.72 s	4,769,023,210 B
896 × 1152	6	1729	183.05 s	4,760,601,834 B

Each run reported two preparation stages and one live preview per step. Renders were visually reviewed. Physical iPhone validation is pending.