LocalMuseAI/mlx-moss-image-realism-stable-yogi-4bit

MOSS-Image Realism by Stable Yogi — MLX INT4 for LocalMuse

Format conversion for the LocalMuse iOS app (profile mossImage5B4Bit) of Realism By Stable Yogi Moss V1.0 by Stable_Yogi, Civitai version 3421295, fp8-scaled file 3315401 (SHA-256 252d4459f19eb041322c725630bc26a0c25df4fc92ac4b7126c753462caf5166). Credit belongs to Stable_Yogi. No additional training or model merge was performed. This does not change or remove the original model permissions.

License and attribution

The author requires attribution and does not allow sharing merges. The Civitai permission flags are preserved in LICENSE-AUTHOR.txt and source.json; read the author's license/addendum. LocalMuse grants no replacement license for these author weights, and this package is not offered under unconditional Apache 2.0 terms.

The text encoder, VAE decoder and tokenizer are the unchanged release components of MOSS-Image-SFT (revision fc5de0d56fc910cbca6f6625a70f5f582b26deb1, OpenMOSS Team, Apache 2.0; Qwen3.5 text encoder by Alibaba Cloud / Qwen Team and FLUX.2 small decoder by Black Forest Labs, both Apache 2.0). Their license is in LICENSE-BASE-APACHE-2.0. This package is not affiliated with or endorsed by Stable_Yogi, the OpenMOSS Team, Alibaba Cloud or Black Forest Labs.

Contents

File	Content
transformer/model.safetensors	the author's MOSS-Image DiT: float8_e4m3fn block projections dequantized with their per-tensor scale to BF16, fused block qkv split into q, k, v, then attention/FFN affine INT4 group 64; embedders, time embedding, modulation, norms and output projection BF16
text_encoder/model.safetensors	Qwen3.5 text model from MOSS-Image-SFT, all linears and the embedding table affine INT4, group 64
vae.safetensors	FP32 FLUX.2 small decoder and latent BatchNorm statistics
tokenizer/	release tokenizer
runtime_config.json, manifest.json	runtime contract and SHA-256 manifest

The checkpoint's own weights carry few-step distillation, so the package has no turbo adapter: runtime_config.json (version 2) sets "embeddedDistillation": true with an empty adapter list.

Settings

The author's settings: 8 steps, CFG 1, Beta schedule. LocalMuse runs Euler over ComfyUI's beta sigmas (alpha = beta = 0.6) on the shift-5 flow grid; the author's ER SDE sampler is stochastic and is not reproduced. Text to Image, batch one, no negative prompt, Hires Fix or Face Detail. Canvases 896 × 1152 (default) or 1024 × 1024: the author's 1024 × 1536 exceeds the validated 4096-token budget.

Validation

Native Swift MLX inference on an Apple M1 Pro / 16 GiB (debug build): 1024 × 1024 / 8 steps / Beta in 209 seconds at 3,570,991,124 peak MLX bytes, 896 × 1152 in 204 seconds. Without any adapter, eight steps give clean, sharp images, confirming the embedded distillation. These are Mac measurements, not iPhone timings.

euler-vs-beta.jpg compares the release Euler schedule with Beta on the same seeds. Conversion: Tools/MOSSImage/convert_moss_image.py --checkpoint Tools/MOSSImage/StableYogiRealism/source.json --transformer-file … in the LocalMuse repository.