DiffusionBear GitHub · Website · X (@DiffusionBear)

MOSS-Image Text Encoder for Apple Silicon: Qwen 3.5 9B (MLX Q8-G64)

Apple MLX 8-bit affine (group_size=64) conversion of the Qwen 3.5 9B text encoder from OpenMOSS-Team/MOSS-Image-SFT. Built for DiffusionBear to run MOSS-Image natively on 16 GB Apple Silicon Macs.

Engineering Summary

Problem: The official BF16 text encoder is ~19 GB. On 16 GB unified memory Macs, running the text encoder alongside the 5B DiT (~10 GB) and VAE exceeds Metal wired limits, forcing slow PyTorch CPU fallback (~61 s/prompt).
Architecture: Hybrid MLX conversion. Heavy MLP projections, full-attention matrices, and embeddings are quantized to 8-bit. All Gated DeltaNet recurrent projections (in_proj_*, out_proj, conv1d, A_log, dt_bias, RMSNorm) remain in bfloat16 to preserve delta-rule state recurrence across 32 layers.
Results: Prompt encoding in 17 s cold / ~4 s warm on Apple GPU (vs 61 s CPU), peak memory 10 GB (vs 19 GB), cosine similarity 0.9995 vs BF16 ground truth.

Sample Generations & Benchmarks (Apple Silicon M1, 16 GB RAM)

All samples generated locally on a MacBook Pro M1 (16 GB Unified Memory) using DiffusionBear and this quantized text encoder.

🏆 Hero Sample: 50-Step Full CFG Trajectory

Prompt: "A cozy corner café at golden hour. A chalkboard sign by the door reads "MOSS Image" in hand-drawn lettering, with “runs on your Mac with DiffusionBear" written smaller underneath. Warm light pours through the window onto a wooden counter with ceramic cups. Shallow depth of field, 35mm photograph, natural colors."

Run Data:
• Hardware: Apple MacBook Pro M1 (16 GB Unified Memory)
• Generation Time: 7,525.2 s (~2 h 05 min)
• LoRA: None (Stock un-distilled model)
• Resolution: 1024 × 1024 · Steps: 50 (Stock CFG Flow-Matching Euler)
• Guidance: 3.0 · Sampler: flow_match_euler_discrete · Seed: 1000

Studio Gallery: DMD Step-Distillation (4/8 Steps) & Virtual Boxes

Tested with community DMD step-distillation adapters (lodestones-8step, darask0-4step) and spatial bounding boxes:

Studio Product: Matte Sneaker (⚡ DMD 8-Step) Travel Poster: Alpine Lakes (⚡ DMD 4-Step)

Prompt: "Studio product photograph of a matte moss-green sneaker resting on a rough grey stone pedestal. Soft diffused lighting, seamless off-white backdrop, the word 'MOSS' embossed on the heel tab."
• Hardware: MacBook Pro M1 (16 GB RAM)
• Generation Time: 825.1 s (~13 min 45 s)
• LoRA: moss_distill/lodestones-8step (scale 1.0)
• Config: 768×768 · 8 steps · CFG 1.0 · moss_dataspace_dmd · Seed: 1003 Prompt: "A minimalist travel poster of a still mountain lake at dawn, flat vector style with soft gradients. A bold title 'ALPINE LAKES' across the top and the tagline 'breathe the quiet' in small capitals."
• Hardware: MacBook Pro M1 (16 GB RAM)
• Generation Time: 464.1 s (~7 min 44 s)
• LoRA: moss_distill/darask0-4step (scale 1.0)
• Config: 512×768 · 4 steps · CFG 1.0 · moss_dataspace_dmd · Seed: 1002

Studio Product: Matte Sneaker (⚡ DMD 8-Step)	Travel Poster: Alpine Lakes (⚡ DMD 4-Step)

Prompt: "Studio product photograph of a matte moss-green sneaker resting on a rough grey stone pedestal. Soft diffused lighting, seamless off-white backdrop, the word 'MOSS' embossed on the heel tab." • Hardware: MacBook Pro M1 (16 GB RAM) • Generation Time: 825.1 s (~13 min 45 s) • LoRA: `moss_distill/lodestones-8step` (scale 1.0) • Config: 768×768 · 8 steps · CFG 1.0 · `moss_dataspace_dmd` · Seed: 1003	Prompt: "A minimalist travel poster of a still mountain lake at dawn, flat vector style with soft gradients. A bold title 'ALPINE LAKES' across the top and the tagline 'breathe the quiet' in small capitals." • Hardware: MacBook Pro M1 (16 GB RAM) • Generation Time: 464.1 s (~7 min 44 s) • LoRA: `moss_distill/darask0-4step` (scale 1.0) • Config: 512×768 · 4 steps · CFG 1.0 · `moss_dataspace_dmd` · Seed: 1002

Botanical Plate: Bryophyta (⚡ DMD 4-Step) Retro Novelty: Shutter Shades (⚡ DMD 4-Step)

Prompt: "A vintage botanical illustration plate of forest moss and ferns on aged cream paper, fine ink linework with watercolor washes. Caption 'Bryophyta', 'plate XII'."
• Hardware: MacBook Pro M1 (16 GB RAM)
• Generation Time: 497.9 s (~8 min 18 s)
• LoRA: moss_distill/darask0-4step (scale 1.0)
• Config: 1024×1024 · 4 steps · CFG 1.0 · moss_dataspace_dmd · Seed: 1005 Prompt: "A studio photograph of a pair of vintage pink plastic shutter shades, resting on a stone pedestal... backdrop with poster, the word 'MOSS x DiffusionBear' engraved on base."
• Hardware: MacBook Pro M1 (16 GB RAM)
• Generation Time: 795.1 s (~13 min 15 s)
• LoRA: moss_distill/darask0-4step (scale 1.0)
• Config: 1024×768 · 4 steps · CFG 1.0 · moss_dataspace_dmd · Seed: 1003

Botanical Plate: Bryophyta (⚡ DMD 4-Step)	Retro Novelty: Shutter Shades (⚡ DMD 4-Step)

Prompt: "A vintage botanical illustration plate of forest moss and ferns on aged cream paper, fine ink linework with watercolor washes. Caption 'Bryophyta', 'plate XII'." • Hardware: MacBook Pro M1 (16 GB RAM) • Generation Time: 497.9 s (~8 min 18 s) • LoRA: `moss_distill/darask0-4step` (scale 1.0) • Config: 1024×1024 · 4 steps · CFG 1.0 · `moss_dataspace_dmd` · Seed: 1005	Prompt: "A studio photograph of a pair of vintage pink plastic shutter shades, resting on a stone pedestal... backdrop with poster, the word 'MOSS x DiffusionBear' engraved on base." • Hardware: MacBook Pro M1 (16 GB RAM) • Generation Time: 795.1 s (~13 min 15 s) • LoRA: `moss_distill/darask0-4step` (scale 1.0) • Config: 1024×768 · 4 steps · CFG 1.0 · `moss_dataspace_dmd` · Seed: 1003

Spatial Virtual Boxes Layout (⚡ DMD 4-Step)

Spatial Virtual Boxes Layout (⚡ DMD 4-Step)

Spatial Layout: Studio scene with shutter shades on pedestal + bounding box `wall_1` (`[35, 0, 527, 1000]`): "Logo from Image 1 posterized on the wall" • Hardware: MacBook Pro M1 (16 GB RAM) • Generation Time: 415.2 s (~6 min 55 s) • LoRA: `moss_distill/darask0-4step` (scale 1.0) • Config: 1024×768 · 4 steps · CFG 1.0 · `moss_dataspace_dmd` · Seed: 1003 • ⚠️ Reference Image Note: An input reference image was provided in this test, but image conditioning does not work on MOSS-Image (MOSS-Image is strictly text & spatial-box conditioned; visual reference inputs are ignored by the model architecture). Layout positioning was driven entirely by the Virtual Box.

Spatial Layout: Studio scene with shutter shades on pedestal + bounding box wall_1 ([35, 0, 527, 1000]): "Logo from Image 1 posterized on the wall"
• Hardware: MacBook Pro M1 (16 GB RAM)
• Generation Time: 415.2 s (~6 min 55 s)
• LoRA: moss_distill/darask0-4step (scale 1.0)
• Config: 1024×768 · 4 steps · CFG 1.0 · moss_dataspace_dmd · Seed: 1003
• ⚠️ Reference Image Note: An input reference image was provided in this test, but image conditioning does not work on MOSS-Image (MOSS-Image is strictly text & spatial-box conditioned; visual reference inputs are ignored by the model architecture). Layout positioning was driven entirely by the Virtual Box.

Benchmark Comparison (Apple M1 16 GB)

Metric	PyTorch BF16 CPU (Upstream)	MLX Q8-G64 + GDN BF16 (This Repo)
Prompt encoding (25 tokens)	~61 s	~17 s (~4 s warm)
Cosine similarity vs. BF16 reference (mean / min)	`1.0000`	`0.99946` / `0.99337`
Text encoder peak Unified Memory	~19 GB	~10 GB
Full 4-step DMD generation (512×512)	463 s	406 s

Note on 4-bit (q4-g64): 4-bit quantization was benchmarked and rejected due to cosine similarity degradation (0.964 mean / 0.642 min) compounding across the 32 recurrent layers. 8-bit maintains cos > 0.99 across all test tokens.

Usage

In DiffusionBear (macOS GUI)

Supported out of the box in DiffusionBear v0.4.0. Selecting MOSS-Image SFT 5B automatically downloads and uses this repository.

Standalone Download

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

加载中…