DiffusionBear — kromoss-mlx-q4

DiffusionBear GitHub  ·  Website  ·  macOS Studio

Kromoss 5B (MLX 4-bit Q4-G64) for Apple Silicon

Apple MLX 4-bit affine (group_size=64) quantization of lodestones/Kromoss (a fine-tune of OpenMOSS-Team/MOSS-Image-SFT), packaged for 16 GB Apple Silicon Macs and DiffusionBear.

Pairs with the shared MLX Q8 text encoder Ouinche/moss-image-qwen3.5-te-mlx-q8g64 so MOSS-Image users do not re-download a 10–19 GB text encoder.

Contents (~4.02 GiB total)

Component File(s) Precision Size
DiT Transformer (5B) transformer/model-q4-g64.safetensors MLX 4-bit (group_size=64) (up from 10.16 GB BF16) 2.66 GiB
8-Step Turbo LoRA 8-steps-dmd2-lora.safetensors Upstream DMD2 rank-128 LoRA (lodestones/Kromoss) 1.11 GiB
FLUX.2 VAE vae/diffusion_pytorch_model.safetensors BF16 (AutoencoderKLFlux2, 4-ch decoder [96,192,384,384]) 238 MiB
Tokenizer tokenizer/* Qwen 3.5 chat-template tokenizer 19 MiB
Scheduler / Meta scheduler/scheduler_config.json, model_index.json FlowMatch (shift=5.0, training_shift=1.0) < 1 MiB

Text Encoder (shared sibling): Ouinche/moss-image-qwen3.5-te-mlx-q8g64 (~10.0 GB). Already present on disk if you installed MOSS-Image SFT in DiffusionBear.


Visual Comparison: Official BF16 (lodestones/Kromoss) vs MLX Q4-G64 (16 GB M1)

Prompt from the official lodestones/Kromoss quick start (1024 × 1024, seed=42):

"a tabby cat holding a sign that says KROMOSS"

Note: Official BF16 references use torch.Generator("cpu").manual_seed(42) while DiffusionBear uses MLX mx.random.seed(42), so initial latent noise differs even on the same seed.

50 steps, CFG 3.0 (uncond_skip_layers=[22]) 8 steps, CFG 1.0 + 8-steps-dmd2-lora
MLX Q4-G64 (this repo, M1 16 GB) MLX Q4 50 steps
4125.85 s total (6.92 s load)
MLX Q4 8 steps DMD2
353.90 s total (6.08 s load)
Official BF16 (lodestones/Kromoss) Official BF16 50 steps
Reference
Official BF16 8 steps DMD2
Reference

Additional Benchmark (MacBook Pro M1, 16 GB Unified Memory)

Kromoss MLX Q4 512x768 8-step botanical plate sample

Prompt: "A vintage botanical illustration plate of glowing bioluminescent moss and emerald ferns on aged cream paper, fine ink linework with watercolor washes, clean typography title KROMOSS at the bottom"

Measured Run (16 GB M1, staged te -> dit -> vae pure-MLX execution):

  • Resolution: 512 × 768 · Steps: 8 · Guidance (CFG): 1.0 · Seed: 1003
  • LoRA: 8-steps-dmd2-lora.safetensors (scale=1.0)
  • Generation Time: 148.65 s (4.76 s model load)

Recommended Settings

  • Fast 8-Step (DMD2 LoRA): 8 steps, CFG 1.0, 8-steps-dmd2-lora.safetensors at scale 1.0 (~354 s at 1024×1024, ~149 s at 512×768 on 16 GB M1).
  • Stock (No LoRA): 50 steps, CFG 3.0, uncond_skip_layers=[22].

Usage

DiffusionBear (macOS GUI)

Supported natively in DiffusionBear: selecting Kromoss 5B downloads this Q4 pack and reuses your existing local qwen3_5-te-q8g64, VAE, and tokenizer when already installed (~3.8 GiB incremental transfer).

Standalone Download

hf download Ouinche/kromoss-mlx-q4 --local-dir ./kromoss
hf download Ouinche/moss-image-qwen3.5-te-mlx-q8g64 --local-dir ./qwen3_5-te-q8g64

Licence & Attribution

This repository is a 4-bit MLX quantized derivative of lodestones/Kromoss.

  • Kromoss weights, 8-step DMD2 LoRA & BF16 reference cat samples (images/cat-*-official.webp): lodestones/Kromoss, Apache-2.0.
  • MOSS-Image architecture: © MOSI.AI / OpenMOSS, OpenMOSS-Team/MOSS-Image-SFT, Apache-2.0.
  • Qwen 3.5 tokenizer: © Alibaba Cloud, Qwen/Qwen3.5-9B, Apache-2.0.
  • MLX 4-bit quantization & packaging: Ouinche / DiffusionBear.

Distributed under the Apache License 2.0. Attribution notices are in NOTICE.

⚖️ License and Attribution

This model is based on the MOSS/KROMOSS architecture developed by the OpenMOSS Team. It is licensed under the Apache License 2.0.

Downloads last month

-

Downloads are not tracked for this model. How to track
MLX
Hardware compatibility
Log In to add your hardware

4-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Ouinche/kromoss-mlx-q4

Quantized
(58)
this model