v0.4.1 · a local-first image generation studio for Apple Silicon
Website · x.com/DiffusionBear · Discord
DiffusionBear A free macOS studio that runs SDXL (and its derivatives), FLUX.1, FLUX.2 Klein 4B/9B, FLUX.3 Klein 5.6B, MOSS-Image SFT 5B, Qwen-Image-2.1 & Turbo, Krea 2 Turbo, and Z-Image-Turbo on your Mac's own Metal GPU. No account, no cloud, no per-image billing.
It is built on Apple's MLX framework, so it is Apple Silicon only. There is no CUDA or CPU path — MLX is macOS/Metal exclusive.
Made by Ouinche — follow on X: x.com/DiffusionBear — released under the MIT licence.
- SDXL Metal Memory Stabilization: Systematic Metal buffer clearing and working-set memory management (+51.9% suite-wide speedup, 4.25x faster on Juggernaut XL Lightning down to 9.2s).
- Generation Pipeline Optimization: Lossless JIT compilation (
@mx.compile), sub-second TAESD VAE decoding, andeuler_trailingtimesteps restoring 100% visual parity (∞ dB PSNR). - Interactive 187-Run Benchmark Arena: Explore all 17 model configurations with our interactive comparison slider at diffusionbear.com/benchmark.html.
- Virtual Boxes on the Studio Canvas: Draw, resize, and label color-coded spatial bounding boxes right on the canvas for FLUX.3-klein 5.6B and MOSS-Image SFT 5B to position subjects, characters, and exact poster typography without ControlNet.
- Smart 16 GB Unified Memory Management:
- Gallery Favorites & Privacy Card Blur:
- Custom Python Samplers:
- DiffusionBee & Draw Things Local Library Import: if you allow DiffusionBear to read theirs directories
DiffusionBear.app is a self-contained macOS application: it bundles the backend, two relocatable Python/MLX runtimes, and the studio interface inside one .app. You do not need to install Python, Node, Xcode, or Homebrew.
⬇ Download DiffusionBear v0.4.1 for macOS (Apple Silicon) · v0.4.1 Release Notes & Checksums
| Detail | |
|---|---|
| Requires | An Apple Silicon Mac (M1 / M2 / M3 / M4 / M5 / M6), macOS 26.2 or later, 16 GB unified memory recommended. |
| Your files | Settings, gallery, LoRA registry, and model store live in ~/Library/Application Support/DiffusionBear. To put the store on an external drive, write an absolute path into ~/Library/Application Support/DiffusionBear/store_path. |
| Model weights | Downloaded on first use from Hugging Face or Civitai into your local store. After that, the app works 100% offline. |
Because DiffusionBear is a free, open-source project signed ad-hoc (without a paid Apple Developer ID notarization certificate), macOS Gatekeeper pauses the very first double-click of a downloaded copy with this dialog:
"DiffusionBear" Not Opened — Apple could not verify "DiffusionBear" is free of malware that may harm your Mac or compromise your privacy.
Follow this one-time 60-second procedure:
- Unzip and move to
/Applications: UnzipDiffusionBear-0.4.1-arm64.zipand dragDiffusionBear.appinto your /Applications folder (do not launch it from~/Downloadsor inside the zip). - Bypass Gatekeeper (choose Option A or Option B):
- Option A — Fastest (Terminal 1-Liner): Open Terminal.app, paste this command, and press Return:
Then double-click
xattr -cr /Applications/DiffusionBear.app
DiffusionBear.appin/Applications— it opens right away! - Option B — No Terminal (System Settings GUI):
- Click Done on the Gatekeeper dialog.
- Open System Settings → Privacy & Security.
- Scroll down to the Security section ("DiffusionBear was blocked to protect your Mac").
- Click Open Anyway, authenticate with Touch ID or your password, and click Open.
- Option A — Fastest (Terminal 1-Liner): Open Terminal.app, paste this command, and press Return:
Once authorized, macOS remembers your choice and never shows the alert again.
Twelve+ checkpoints across eight families, 4-bit quantised and memory-tuned for 16 GB of unified memory. Switching engines swaps weights on demand rather than keeping everything resident.
| Model | Family | Default steps | What it supports |
|---|---|---|---|
| SDXL & Derivatives (Juggernaut XL Lightning, RealVisXL V5.0 Lightning / Hyper-SD, Juggernaut XI v11, Custom SDXL / Pony / Illustrious) | SDXL | 4–50 | Ultra-fast 4-step Lightning (~9s) or full 20–50 step custom .safetensors checkpoints (auto-converted), negative prompts, sampler picker, CLIP Skip, DeepCache, and TAESD fast decode. |
| FLUX.1-schnell | FLUX.1 | 4 | Classic 4-step rectified flow, 1 reference image, FLUX.1 LoRAs, and fast-VAE decode. |
| FLUX.2-klein 4B | FLUX.2 | 4 | Up to 10 reference images, FLUX.2 LoRAs, and generative fill. Guidance-distilled (no negative prompt). |
| FLUX.2-klein 9B | FLUX.2 | 4 | 9-billion-parameter flagship DiT with up to 10 reference images, LoRAs, and superior detail. |
| FLUX.3-klein 5.6B | FLUX.3 | 16 | 16-step flow-matching distilled, up to 10 reference images, and interactive Virtual Boxes on the Studio Canvas. |
| MOSS-Image SFT 5B (+ Base 5B) | MOSS | 4–8 (DMD) / 25 | Interactive Virtual Boxes, reference image, LoRAs, MLX q8-g64 text encoder, and automatic 4-step (⚡ DMD 4) / 8-step (⚡ DMD 8) distillation adapters. |
| Qwen-Image-2.1 | Qwen-Image | 25 | 7B single-stream DiT with Qwen3-VL text encoder, 1 reference image, and negative prompts. |
| Qwen-Image-Turbo-2.1 | Qwen-Image | 8 | 8-step distilled Qwen-Image 2.1 variant for fast generation on 16 GB Macs. |
| Krea 2 Turbo 13B | Krea 2 | 4–8 | 13B photorealistic engine with 1 reference image, LoRAs, persistent Q4 text-encoder cache, and auto-applied 4-step distilled LoRA. |
| Z-Image-Turbo 6B | Z-Image | 6–8 | Compact 6B engine with 1 reference image, LoRAs, fast VAE, and crisp architectural geometry. |
Bring any SDXL, Pony, or Illustrious checkpoint you own: single-file Civitai .safetensors checkpoints are converted in place, unlisted models can be onboarded from a local folder or your Hugging Face cache, and the app also browses models and LoRAs already downloaded by Draw Things and DiffusionBee (strictly read-only).
All images below were generated locally in DiffusionBear on an Apple MacBook Pro M1 with 16 GB of RAM.
☕ 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."
Generation Metadata (Generated Locally in DiffusionBear):
• Hardware: Apple MacBook Pro M1 (16 GB Unified Memory / RAM)
• Generation Time:7,525.2 s(~2 h 05 min — full stock 50-step CFG trajectory)
• Active LoRAs: None (Base MOSS-Image pipeline)
• Model:OpenMOSS-Team/MOSS-Image-SFT(5B Parameters, Native MLX Text Encoder)
• Resolution:1024 × 1024· Steps:50(Stock CFG Flow-Matching Euler Trajectory)
• Guidance (CFG):3.0· Sampler:flow_match_euler_discrete· Seed:1000
• Text Encoder: Custom 8-bit MLX Quantized Qwen3.5 TE (Ouinche/moss-image-qwen3.5-te-mlx-q8g64)
The following original images demonstrate the versatility of MOSS-Image SFT 5B inside DiffusionBear on a MacBook Pro M1 (16 GB RAM), using both built-in DMD step-distillation adapters (generating in 7–13 minutes instead of 2 hours) and interactive Virtual Boxes spatial layout:
| Studio Product — Matte Sneaker (⚡ DMD 8-Step) | Alpine Lakes Minimalist Poster (⚡ 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. Crisp fabric texture, subtle shadow, high detail." • Hardware: MacBook Pro M1 (16 GB RAM) • Generation time: 825.1 s (~13 min 45 s) • Active LoRA: lodestones-8step (scale 1.0, ⚡ DMD 8-step kit)• Model: MOSS-Image SFT 5B + ⚡ DMD 8 • Size: 768×768 · Steps: 8 (CFG 1.0) • Sampler: 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 along the bottom. Teal, cream and dusty pink palette, clean typography, centered composition." • Hardware: MacBook Pro M1 (16 GB RAM) • Generation time: 464.1 s (~7 min 44 s) • Active LoRA: darask0-4step (scale 1.0, ⚡ DMD 4-step kit)• Model: MOSS-Image SFT 5B + ⚡ DMD 4 • Size: 512×768 · Steps: 4 (CFG 1.0) • Sampler: moss_dataspace_dmd · Seed: 1002 |
| Vintage Botanical Plate — Bryophyta (⚡ DMD 4-Step) | Vintage Pink 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. A caption at the bottom in serif type reads 'Bryophyta' and below it, smaller, 'plate XII'. Scientific illustration style, balanced layout." • Hardware: MacBook Pro M1 (16 GB RAM) • Generation time: 497.9 s (~8 min 18 s) • Active LoRA: darask0-4step (scale 1.0, ⚡ DMD 4-step kit)• Model: MOSS-Image SFT 5B + ⚡ DMD 4 • Size: 1024×1024 · Steps: 4 (CFG 1.0) • Sampler: moss_dataspace_dmd · Seed: 1005 |
Prompt: "A studio photograph of a pair of vintage pink plastic shutter shades, resting on a stone pedestal. Facing the viewer. Novelty sunglasses feature a molded pink frame with horizontal slats... backdrop with poster on the wall, the word 'MOSS x DiffusionBear' engraved on the stone base." • Hardware: MacBook Pro M1 (16 GB RAM) • Generation time: 795.1 s (~13 min 15 s) • Active LoRA: darask0-4step (scale 1.0, ⚡ DMD 4-step kit)• Model: MOSS-Image SFT 5B + ⚡ DMD 4 • Size: 1024×768 · Steps: 4 (CFG 1.0) • Sampler: moss_dataspace_dmd · Seed: 1003 |
| Interactive Virtual Boxes: Multi-Subject Spatial Layout (⚡ DMD 4-Step) |
|---|
 |
Virtual Box Spatial Layout: Studio scene with shutter shades on stone pedestal + interactive bounded 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) • Active LoRA: darask0-4step (scale 1.0, ⚡ DMD 4-step kit)• Model: MOSS-Image SFT 5B + ⚡ DMD 4 ( darask0-4step) with Studio Virtual Boxes• Size: 1024×768 · Steps: 4 (CFG 1.0) · Sampler: moss_dataspace_dmd · Seed: 1003• |
| Studio Generation (Dark Mode) | Studio Generation (Light Mode) |
|---|---|
 |
 |
| Gallery / Browser (Dark Mode) | Gallery / Browser (Light Mode) |
|---|---|
 |
 |
| Models: Installed & Manage | Models: External Disk Detection (Draw Things & DiffusionBee) |
|---|---|
 |
 |
| Parameters: Engine, GPU Memory & Model Defaults | Parameters: Language & System Preferences |
|---|---|
 |
 |
Three engines (FLUX.3-klein, MOSS-Image SFT and Base) accept spatial boxes: instead of hoping the model puts the cat in the right corner, you draw a box on the Studio Canvas and say what belongs inside it. The canvas can show a reference image as a backdrop, so you compose over the actual picture you want.
- SPATIAL BOX panel: presets (Center / Left / Right / Top / Bottom / full frame), free boxes you drag on the canvas, and a live count.
- Compose turns your caption + boxes into the layout description the engine understands; JSON shows exactly what will be sent.
- Reference backdrops: overlay any gallery image or reference behind the boxes, hide it or remove it at will.
- The layout travels with the generation metadata, so reusing parameters reproduces the composition.
For FLUX.2-klein, the equivalent is generative fill: paint a mask on a gallery image and regenerate just that region.
Running 4–13 B parameter models on a 16 GB Mac is a memory-budget problem. The app manages it; you watch (Parameters → Engine & GPU) rather than tune:
- Auto wired-limit. Each engine's GPU wired budget is derived from your
machine —
min(68% of unified memory, Apple's recommended working set), about 10.9 GB on a 16 GB Mac. An explicit number you set still wins;0means unbounded. - Idle auto-release. Resident pipelines (a full FLUX/Krea2 pipeline is
~10 GB) are released after 5 minutes of inactivity (300 s default,
tunable per engine;
0= never release). Reloading afterwards costs a few seconds, not minutes — macOS keeps the weight files in its file cache. - Quantised everything. Weights are 4-bit by default; the MOSS text encoder is swapped for the app's own MLX q8g64 build (~2.2 GB) and the Krea 2 encoder is quantised once and persisted, so later sessions skip both the 7.5 GB bf16 load and the quantisation pass.
- Live panel. Metal memory in use, peak, the wired limits, which pipelines are resident, and the idle countdowns — all visible, nothing hidden.
- Storage cleanup. The Parameters tab reports generated-image size, TAESD state and settings, and the Models tab clears per-repo Hugging Face cache entries to reclaim disk space.
Each MOSS DMD kit is three files that only work together: a projection LoRA, a distilled fp32 time embedding, and a dataspace file holding the CFG-free step schedule. The DMD distillation control has four positions:
| Mode | Adapter used | Schedule |
|---|---|---|
| Auto (default) | follows the step slider: ≤4 steps → DMD 4, ≤8 → DMD 8, above → stock | stock CFG trajectory above 8 steps |
| ⚡ DMD 4 | darask0's 4-step kit (moss_distill/darask0-4step) |
CFG-free, engine clamps the run to the 4 codes |
| ⚡ DMD 8 | lodestones' 8-step kit (lodestones/MOSS-Image-SFT-8-steps-dmd, Apache-2.0) |
CFG-free, engine clamps the run to the 8 codes |
| Stock | none | full CFG trajectory at the requested steps (~20–45 min) |
Forcing DMD 4 or DMD 8 at a higher requested step count still runs the 4- or 8-step codes (the engine clamps the loop and drops CFG, which has no effect on the dataspace schedule). On a 16 GB M1 that is roughly 1 min (4 steps) / 3 min (8 steps) at 768×512, against 20–45 min stock. If a kit is missing the app downloads it automatically when the run needs it (resumable); with no network it falls back to the stock trajectory rather than failing.
- Generation runs on the Metal GPU through MLX (
mfluxfor the flow-matching families, a native MLX daemon for SDXL, dedicated engines for MOSS and Qwen). Weights are 4-bit so a 4–13 B model fits a 16 GB machine. - The prompt enhancer is a local LLM too (Qwen2.5 0.5B or 1.5B, MLX) — rewriting a prompt never leaves the machine.
- No telemetry, no analytics, no network calls at generation time. Once the weights are cached the app works fully offline.
- Tokens (Civitai API key, Hugging Face token — both optional, both only for smoother downloads of gated or rate-limited repos) are stored on your disk and sent only to the service they belong to.
Unified memory is why this works at all on a 16 GB machine: the CPU and GPU share one pool, so a quantised model stays resident with no CPU↔GPU copying — exactly what MLX is built to exploit.
Measured on a 2021 MacBook Pro M1 (16 GB) — the reference machine this project is tuned for. Comprehensive v0.4.1 snapshot (187 timed studio generations across 17 model configurations at 512×768 resolution):
| Model | Steps | Median | Range | Family / Notes |
|---|---|---|---|---|
| Juggernaut XL Lightning | 4 | 9.2 s | 6.4 – 10.6 s | SDXL · 4.25x faster (+76.5%), TAESD fast VAE |
| Juggernaut XI v11 | 8 | 42.5 s | 29.7 – 46.9 s | SDXL · Hyper-SD 8-step, 2.15x faster |
| RealVisXL V5.0 | 8 | 43.3 s | 27.6 – 50.0 s | SDXL · Hyper-SD 8-step, 3.63x faster |
| FLUX.2-klein 4B | 4 | 48.3 s | 35.1 – 49.2 s | FLUX.2 · Compact 4B rectified flow |
| FLUX.3-klein 5.6B | 4 | 54.8 s | 35.5 – 56.8 s | FLUX.3 · 5.6B flow-matching + Virtual Boxes |
| FLUX.1-schnell | 4 | 1:20 min | 55.8 – 82.4 s | FLUX.1 · Classic 4-step rectified flow |
| FLUX.2-klein 4B (8s) | 8 | 1:36 min | 70.8 – 96.6 s | FLUX.2 · High-detail 8-step mode |
| FLUX.2-klein 9B | 4 | 1:45 min | 76.8 – 116.8 s | FLUX.2 · Flagship 9B rectified flow |
| FLUX.3-klein 5.6B (8s) | 8 | 1:46 min | 70.6 – 110.5 s | FLUX.3 · High-detail 8-step mode |
| Pony Diffusion V6 XL | 25 | 1:50 min | 1:08 – 1:59 min | SDXL · Full 25-step anime/illustrative |
| Z-Image Turbo 6B | 8 | 2:04 min | 1:17 – 2:07 min | Z-Image · Sharp geometry 8-step |
| Krea 2 Turbo 13B | 4 | 2:55 min | 2:33 – 3:16 min | Krea 2 · 13B photorealistic (Distilled) |
| Qwen-Image 2.1 Turbo | 8 | 3:20 min | 2:24 – 3:47 min | Qwen-Image · 8-step distilled 7B DiT |
| FLUX.2-klein 9B (8s) | 8 | 3:29 min | 2:35 – 3:30 min | FLUX.2 · Flagship 9B 8-step |
| Krea 2 Turbo 13B (8s) | 8 | 5:14 min | 3:38 – 5:55 min | Krea 2 · 13B native 8-step |
| Qwen-Image 2.1 | 20 | 7:26 min | 5:12 – 7:44 min | Qwen-Image · Full 20-step 7B DiT |
| MOSS-Image SFT 5B | 25 | 26:28 min | 20:19 – 27:45 min | MOSS · Full 25s stock (~1 min with ⚡ DMD 4) |
For the interactive side-by-side visual duel arena, prompt latency breakdowns, and historical September comparisons, visit diffusionbear.com/benchmark.html.
Head-to-head comparisons on original, royalty-free scenes. Explore interactive side-by-side comparison sliders on diffusionbear.com/benchmark.html or view our video write-up at ouinche.com.
- Model browser & downloader — search Hugging Face and Civitai side by side, classify, download with progress/speed/cancel, direct URLs, and convert single-file Civitai SDXL checkpoints to runnable diffusers directories out-of-process.
- Reference images — FLUX.2/FLUX.3 take up to 10 (
Image 1,Image 2, …), Z-Image, Krea 2, Qwen and FLUX.1 take one; studio-canvas backdrops on the box-capable engines. - LoRA manager — drag a
.safetensorsin or import from Civitai; FLUX.1, FLUX.2, FLUX.3, Z-Image, Krea 2 and SDXL (rank-concat stacking) formats; trigger words become clickable chips; a LoRA that does not fit the selected model is refused before generating, with the mismatch named. - Prompt enhancer — local Qwen2.5 0.5B/1.5B rewrites your prompt per engine family, prose or JSON, with editable system prompts and trigger-word preservation.
- Custom samplers — from a local
.pyfile or a URL; protected built-in default; a broken custom sampler falls back to the engine default, never to a failed generation. - Queue & recovery — one generation at a time behind a FIFO worker; a job interrupted by a crash or quit is recoverable and re-queueable.
- Gallery — tags, search, favourites, one-click Use-as-Reference, 2×/4× Lanczos upscaling, Civitai-compliant PNG metadata (with a stealth mode that omits it all).
- Interface in English, French, German, Italian, Spanish, Simplified Chinese, Japanese, Portuguese and Korean.
- ⚖ Licences tab — the licence of every bundled package and every supported model, in one place.
Everything above is the app; this section is the minimum needed to build and run it from source. An Apple Silicon Mac, Python 3.10 + 3.14 and Node.js 18+ are required.
Build the app (three scripts, in order — output lands in
packaging/dist/DiffusionBear.app; INSTALL_APP=1 also installs it to
/Applications):
packaging/build_runtime_venv.sh # main runtime: Python 3.10 + mflux
packaging/build_sdxl_venv.sh # SDXL engine runtime: Python 3.14 + mlx_diffuser
packaging/build_app.sh # assembles and ad-hoc signs DiffusionBear.appThe scripts audit their own output and fail the build on absolute build paths,
unrelocatable interpreters, weights or credential-shaped files. Install with the
build script's own install step — a hand cp -R of a signed bundle is known to
silently skip files.
Run from a checkout (development):
python3 -m venv venv && python3 -m venv venv-sdxl
./venv/bin/python -m pip install -r backend/requirements.txt
./venv-sdxl/bin/python -m pip install -r backend/requirements-sdxl.txt
cd frontend && npm install && cd ..
./run.sh # backend on 8001, Vite on 5174, opens the UIDev mode is ./dev-backend.sh plus cd frontend && npm run dev; lint and build
are npm run lint and npm run build. Backend tests run with
./venv/bin/python backend/test_<name>.py. Ports 8001/5174 are deliberate — do
not "fix" them.
Public-source export. packaging/export_public.sh builds a filtered tree
(internal docs, tests, scratch excluded), audits it for local paths,
credential-shaped files and weights, and refuses to ship anything unsafe;
--push publishes to the public repository.
- CHANGELOG.md — what changed in each release
- RELEASE_NOTES.md — release notes and the Gatekeeper guide
- CONTRIBUTING.md — how to build, verify and propose a change
- SECURITY.md — privacy posture and vulnerability reporting
- USER_GUIDE.md — full usage guide
- docs/VirtualBox.md — How does Virtual Boxes works ?
DiffusionBear is written and maintained by Ouinche (https://github.com/OuincheWinch). Follow on X: x.com/DiffusionBear.
DiffusionBear is heavily co-coded by AI — LLM do a large share of the writing. That is a deliberate choice, not an accident.
Please do not open a public issue for a security problem. Use GitHub's
private vulnerability report,
or email the author -hello -at- -DiffusionBear-.-com- with the subject [DiffusionBear security].
Scope and the privacy posture are in SECURITY.md.
MIT — Copyright © 2026 Ouinche. This covers the project's source code only. Model weights carry their own terms (FLUX.2/FLUX.3 are Black Forest Labs non-commercial; the Juggernaut and RealVisXL checkpoints are non-commercial too), and none of them are redistributed here — they are downloaded from Hugging Face or Civitai on first use. The app shows the full per-package and per-model licences in its ⚖ Licences tab.