🐻 DiffusionBear Research • Apple Silicon Laboratory

MOSS-Image 5B on Apple Silicon: SFT vs DMD Distillation Benchmark

👤 By Research Team 📅 October 19, 2026 💻 Hardware: Apple Silicon M1 16GB Unified Memory 🧠 Text encoder: Qwen 3.5 9B — MLX 8-bit (first GPU MLX runtime)

🧩 0. Why MOSS-Image needed a custom pipeline on the Mac

The official MOSS-Image release is a PyTorch/diffusers pipeline written for big NVIDIA GPUs. As shipped, it wants a ~19 GB bf16 Qwen 3.5 9B text encoder, a ~10 GB 5B DiT and an fp32 VAE resident at the same time — about 30 GB, roughly twice what a 16 GB Mac has. A naive port either crashes or falls back to the CPU for the text encoder (~61 s per prompt) and then takes 40+ minutes per image.

DiffusionBear v0.4.1 rebuilds that pipeline stage by stage so it fits a 16 GB M1 and finishes in minutes:

1Virtual Boxes → layout JSONThe boxes you draw on the Studio Canvas are serialized into MOSS's native summary/items layout (abs_centralized, yxyx): background, wordmark, logo slot, captions.0 GB · in-app

2MLX 8-bit text encoderOur own conversion of the exact OpenMOSS Qwen 3.5 9B weights to MLX q8-g64, with Gated-DeltaNet layers kept bf16. Runs on the GPU: ~17 s vs ~61 s on CPU, parity cos 0.9995.~10 GB peak, then freed

3Embedding cache & hand-offThe prompt embedding is cached on disk, the encoder is dropped and the Metal pool cleared before the DiT is loaded — the two giants never share memory.−10 GB before denoise

4DiT + DMD distillationAuto mode picks the adapter from your step count: ≤4 → ⚡ DMD 4 (darask0), ≤8 → ⚡ DMD 8 (lodestones), above → stock CFG. DMD kits fold CFG away: 1 forward per step instead of 2.~10 GB DiT

5Tiled VAE decodeFLUX.2-family VAE decoded in 512 px tiles with cosine blending, so a 1024² poster does not spike memory at the very last second.−4.5 GB at 1024²

DMD kits are downloaded on first use (~1.2 GB each, resumable): a projection LoRA, a distilled fp32 time embedding and a dataspace schedule. With no network, the app falls back to the stock trajectory instead of failing.
Why not a ready-made 4-bit Qwen? Community conversions quantise their reference weights; MOSS uses a bare Qwen3_5Model subset. Converting locally guarantees the 8-bit scales come from the exact tensors MOSS reads — and lets us keep the delta-rule projections in bf16. 4-bit was measured and rejected (parity fell to 0.964).
Nothing to configure. In the Generate tab, pick MOSS-Image SFT 5B, draw your boxes, and leave the DMD distillation control on Auto.

      Result on a 16 GB M1: the same 768×512 poster goes from ~43 min (stock, 30 steps) to ~13 min (DMD 8) or ~7.5 min (DMD 4) — compare all three below.
    

🐻 1. One prompt, one seed, three inference pipelines

A single press-kit poster — a structured layout (virtual bounding boxes: alpine-lake background, wortmark, logo slot, version line, engine credits) — was rendered by the same MOSS-Image SFT 5B checkpoint through the three pipelines DiffusionBear supports. Identical seed (1024), identical resolution (768×512), identical layout JSON, identical text-encoder path. Only the sampler schedule and the CFG handling change:

Stock SFT (30 steps, CFG 3.0): the reference pipeline. Flow-matching Euler over a discrete schedule, two transformer forwards per step (conditional + unconditional). Maximum detail budget, maximum wall time (~43–46 min per poster).
DMD 8-step (lodestones adapter): the Distribution Matching Distillation adapter replaces the sampling schedule with 8 learned dataspace codes and folds the t=1 CFG vector into a fp32 time embedder — one forward per step, no unconditional pass (~13–14 min).
DMD 4-step (darask0 adapter): the same machinery at 4 codes: fastest MOSS arm (~7.5–8.5 min), the most compressed trajectory.

Distillation adapters do not merely skip steps: each arm is a genuinely different sampler (different trajectory family, CFG-free vs CFG-on), like our earlier Krea-2 4/6/8-step finding. All comparisons below must be read as pipeline comparisons, not quality gradations of one pipeline.

2577s

Stock SFT 30 · CFG 3.0

784s

DMD 8-step · CFG-free

449s

DMD 4-step · CFG-free

⚡ Visual Comparator (drag the divider)

Structured layout • 768×512 Seed: 1024

Loading…

Duel: pick any two pipelines MOSS-Image SFT 5B

LEFT: RIGHT:

Right

Left

⮜ ⮞

Gen time (L / R):-

Speedup (stock / arm):-

💡 The two styles were generated with identical layout, seed and boxes — only the style description on the background box differs. Drag to compare any two arms side by side.

Pipeline	Adapter / LoRA	Sampler	Steps	CFG	Transformer / step	Vectorial dawn	Belle Époque 1900s	s/step (avg)
Stock SFT	—	flow_match_euler_discrete	30	3.0 (+ negative slot)	2 (cond + uncond)	2577.3s	2738.8s	≈85.8 / 91.3
DMD 8-step	lodestones-8step	moss_dataspace_dmd	8	1.0 (folded)	1	784.0s	859.2s	≈98.0 / 107.4
DMD 4-step	darask0-4step	moss_dataspace_dmd	4	1.0 (folded)	1	449.4s	502.9s	≈112.4 / 125.7*

🔬 2. What every arm shared

Checkpoint: OpenMOSS-Team/MOSS-Image-SFT 5B (DiT), bf16, MLX.
Text encoder: Qwen 3.5 9B running on GPU through our MLX 8-bit conversion (g64, GDN layers kept bf16) — parity 0.9995 cos vs the torch bf16 reference (full write-up), ~17s per prompt instead of ~61s on CPU.
Structured layout: virtual bounding boxes serialized into the official summary/items JSON layout (abs_centralized, yxyx): full-surface background box, bold wordmark, logo slot, "V 0.4.1 beta" + "now runs", engine credits in the bottom corners.
Seed 1024, 768×512, FLUX.2-family fp32 VAE decode, no fast-VAE.
*s/step caveat: DMD arms carry ~2-3 min of fixed overhead (text-encoder pass, 252 LoRA wraps, adapter load) that dilutes their per-step average; the stock arm pays two forwards per step instead.

🚀 Practical takeaways

In-app, DiffusionBear auto-selects the DMD adapters for runs ≤ 8 steps (4 → darask0, 5–8 → lodestones) and keeps stock trajectories above. For poster work where text on the layout matters most, the 30-step stock arm renders the most stable lettering; for fast iteration the 4-step arm is 5.7× cheaper. All arms run 100% locally, offline, on a single 16 GB M1 — no cloud, no subscription.