A native macOS image generator for FLUX.2 Klein, running fully on-device on Apple Silicon through MLX — no Python, no ComfyUI, no local server. Type a prompt in any language, press Generate, get an image. Reference images (image-to-image), LoRA adapters — and since 2.0 a full on-device LoRA Training Studio, a 247-scene Library, and a gallery where every image carries its own reproducible recipe.
Not affiliated with SCB 10X's "Typhoon" LLMs or any other Typhoon-named project. "FLUX" is a trademark of Black Forest Labs; this is an independent client for it.
The Generate tab: model tier & quality, a real steps slider, aspect presets, references, live preview — with full system telemetry on the right, all rendering fully on-device.
- Apple Silicon Mac (M1 or newer). Intel Macs are not supported (MLX is Apple-Silicon only).
- macOS 14 (Sonoma) or newer.
- RAM: 16 GB is comfortable for Klein 4B; Klein 9B wants 32 GB (it peaks ~19–20 GB).
The app picks a tier to match your RAM and warns before anything risks swapping.
- LoRA training: Klein 4B trains on a 32 GB Mac (peaks ~22 GB at 512 px); Klein 9B training needs 48 GB+ — the app is honest about it and refuses a run that wouldn't fit.
- 8 GB is not recommended. The app runs (Klein 4B only, with output resolution auto-capped to ~896 px), but it leans on swap and is slow (~2 min/image), and memory-heavier paths — AI prompt enhancement or combining several reference images at once — can run out of memory. It won't harm your Mac, but don't expect comfortable work.
- Disk: ~8 GB for Klein 4B (transformer + encoder), ~26 GB for Klein 9B.
The app downloads model weights itself the first time you need them; nothing is bundled in this repo. You accept each model's license under your own Hugging Face account.
| Tier | Params | Text encoder | License | Download |
|---|---|---|---|---|
| Klein 4B (default) | 4B | Qwen3-4B | Apache-2.0 | Tokenless — just press Download |
| Klein 9B | 9B | Qwen3-8B | FLUX.2 Klein [Non-Commercial] License | Gated — needs a free HF token + license acceptance |
Both tiers share one VAE. Guidance is fixed at 1.0 (the distilled Klein tiers are CFG-free); steps default to Klein's distilled sweet spot of 4 and are adjustable in the app. The Qwen3 encoder understands prompts in any language natively — no translation needed.
For the gated Klein 9B, create a free Read token at huggingface.co/settings/tokens, accept the Klein 9B license (the app links you straight to it), and paste the token into the app's Models tab. Klein 4B needs no token.
-
Download
Typhoonminigen.appfrom the Releases page. -
macOS quarantines apps downloaded from the internet. Because this build isn't signed with a paid Apple Developer certificate yet, you have to clear that flag once. This is the standard way to run an unsigned open-source app — not a hack:
Easiest — Terminal (one line):
xattr -d com.apple.quarantine /path/to/Typhoonminigen.app
(Drag the app onto the Terminal window to paste its path.)
Or — without Terminal: double-click the app, let macOS refuse it, then open System Settings → Privacy & Security, scroll down, and click “Open Anyway” next to the Typhoonminigen message. Confirm once.
-
Open the app, go to the Models tab, and download a tier.
Building it yourself? You don't need any of the above — locally built apps aren't quarantined.
Requires Xcode 26 with the Metal Toolchain (xcodebuild -downloadComponent MetalToolchain,
one-time, ~688 MB).
Build with
xcodebuild, notswift build. Plain SwiftPM can't compile MLX's Metal shaders, and the app would crash at runtime withFailed to load the default metallib.
The engine is temporarily a local fork: upstream
flux-2-swift-mlx v2.4.0 plus vendored
patches (patches/) — a RoPE-cache fix (corrupted renders whose size transposes the
previous one, e.g. portrait right after landscape) and a LoRA-training DOP caption fix.
Package.swift expects it as a sibling checkout named flux-2-swift-mlx-patched; once upstream
ships the fixes, the plain URL pin comes back and this step goes away.
git clone https://github.com/abgitdev/Typhoonminigen.git
# Engine fork: clone upstream v2.4.0 NEXT TO the app checkout, apply the vendored patches (in order)
git clone --branch v2.4.0 --depth 1 https://github.com/VincentGourbin/flux-2-swift-mlx.git flux-2-swift-mlx-patched
for p in Typhoonminigen/patches/*.patch; do git -C flux-2-swift-mlx-patched apply "../$p"; done
cd Typhoonminigen
tools/bundle_app.sh Release # builds + assembles ~/Desktop/Typhoonminigen.appThe first build compiles MLX (including Metal shaders) from source — that part is slow. The
dependency mlx-swift is pinned to 0.30.6 (0.31.4 breaks the engine's training code).
- Text-to-image in any language.
- Image-to-image — up to 4 reference images, optionally described for you by a small on-device vision model (Qwen3.5-VLM 4B — a separate ~3 GB download the first time you use Describe, no token).
- LoRA Training Studio — train your own LoRA fully on this Mac: photos in, auto-captions, live loss curve, checkpoint comparison, publish to your library. No cloud, no Python.
- LoRA adapters — import
.safetensors(or straight from a Hugging Face URL), automatic tier compatibility check, per-adapter trigger words, any number fused at once. - Scene Library — 247 ready-made, hand-tuned scenes across 19 studios, each previewed by a real render of that exact recipe — and fully editable: hide anything, add your own cards.
- Prompt presets — curated, BFL-ordered phrase chips (camera, lighting, look…) plus your own.
- Queue & batch — render a series; you're notified when a long run finishes.
- Gallery — every image keeps its full recipe inside the PNG; drop one back on the canvas (or use “Remix”) to restore the exact settings, or drag any image straight out to Finder.
- Upscale ×2/×4 via Real-ESRGAN.
- Live system telemetry — CPU / GPU / RAM / MLX memory / disk.
- Model management — download, import existing MLX weights, delete, unload to reclaim RAM.
What it doesn't do — Klein is a 4-step distilled model on a pinned engine: no negative prompts, no sampler/CFG choice, and no inpainting, outpainting, masks or ControlNet. Image-to-image regenerates the whole scene from your references rather than editing part of one.
Version 2.0 is the biggest release so far: 57 new features across seven areas, plus an engine-level render fix and 150+ smaller fixes from two full audits of the codebase.
- Workshops — collect 10–50 photos of one concept, pick a trigger word, edit captions per
image; sidecar
.txtcaptions import automatically. - Auto-caption with the local vision model — runs through the training queue, every caption saved the moment it's written.
- Editable captioning instruction per workshop, with presets per concept type (person, pet, object, style).
- Bulk caption tools — find & replace with a live “affects N of M” preview, append/prepend.
- Trigger-in-caption lint with a one-click fix (the engine adds the trigger at train time).
- Training forge — tier, steps, resolution, rank, checkpoint cadence, regularization; live readiness gates (weights on disk? RAM fits?); the dataset is frozen immutably per run.
- Dataset insight — a live “each of your N images is seen ~X×” line with sourced advice bands while you pick the step count.
- Live training monitor — loss curve (raw + smoothed), checkpoint markers, Pause, mid-run Checkpoint, and a graceful Stop that saves a checkpoint first.
- Training queue — line up several trainings; they run one after another, and renders queued during a training fire the moment the trainer goes idle.
- Checkpoint compare matrix — validation prompts × checkpoints, with the engine's no-LoRA baseline pinned on top; a lightbox scrubs one prompt through training time on the same seed.
- Publish the best checkpoint straight into your LoRA library — trigger word set, with an auto-generated, exportable model card.
- Extend a finished training (+250/+500/+1000 steps) from its last checkpoint with optimizer momentum intact — the fix for “looks undertrained”.
- Clone any training as a draft to A/B a single setting.
- Reusable validation prompt sets, with an optional pinned seed per prompt.
- Rolling checkpoint retention (keep the last N weight sets — previews always stay) and mirror augmentation to double a small dataset at forge time.
- Honest memory gates — Klein 4B trains on a 32 GB Mac; a 9B run that wouldn't fit is refused up front instead of freezing your Mac.
- Unlimited stacking (was 2) — add any number of compatible adapters, each with its own strength.
- Strength sweep — one click queues the same prompt + seed at 0.6 → 1.1 to find the sweet spot.
- Pin a default strength per adapter — pre-selected on every future add.
- Bake a strength into a copy — so 1.0 of the new file means exactly your 0.8.
- Merge several same-tier adapters into one
.safetensors. - Identity cards — description, default strength, provenance and date for every adapter.
- Import from a Hugging Face URL — compatibility is checked against the file header before the big download starts.
- Metadata round-trip — trigger and base-model tags are read from ai-toolkit/kohya metadata on import and written back on publish, so adapters stay portable.
- Copy-trigger and reveal-in-Finder buttons on every adapter row.
[trigger]placeholder — drop the active LoRA's trigger word exactly where you want it in a prompt.
- Real steps slider (1–50) — Klein's distilled sweet spot of 4 stays the default; raise it to chase detail, no model reload.
- Quality presets — pick the quantization per tier: Klein 4B Low/Medium/Better/Best, Klein 9B Standard/Best (with an honest note when two rungs render identically).
- Seven aspect-ratio presets, all at Klein's ~1 MP sweet spot.
- Up to 4 reference images (was 3).
- Enhance-prompt reveal — the exact rewritten prompt the model saw is saved with the image and shown in the gallery. No more black box.
- AI insights: Ask — ask the local vision model anything about your render.
- AI insights: Compare — score how closely the result matches your reference (scene + style, 0–10).
- Reference ⇄ result flip — click the reference thumbnail on a render to flip between them in place; the honest identity check.
- Save as card — snapshot the whole form (prompt, chips, model, steps, quality, size, seed) plus the on-screen render into your own Library card.
- Live render telemetry — the canvas says what's happening before step 1, shows s/step during the render, and every image records its speed and true peak RAM.
- Rebuilt from scratch: 247 ready scenes across 19 studios — cinema, portrait, product, interiors, exteriors, food, automotive, anime, game concept art and more.
- Every card's preview is a real render of that exact recipe — not an illustration.
- Cards restore their pinned settings (model, steps, quality, size — many pin the exact seed, so rendering unchanged reproduces the preview).
- Fully editable — hide any built-in card or whole studio, Delete all, Restore all; nothing is ever destroyed.
- Your own cards and sections — build cards in an editor with cover, chips and pinned settings, or save one straight from Generate.
- Rebuilt Models tab — grouped per tier, one row per component, including the base (non-distilled) weights LoRA training runs on.
- Resumable downloads — interrupted, cancelled, even crashed downloads continue from where they stopped.
- Parallel downloads — big files stream as six simultaneous range chunks; every segment resumes from its exact offset.
- Download queue — line up several downloads; they run one after another.
- The Models tab stays usable during a download — queue more, delete other models.
- The vision model is visible and pre-downloadable — 4-bit (recommended) and 8-bit variants, with an explicit “active” switch.
- Search & filters — search by prompt, LoRA name or exact seed; filter by model, adapter or quality, with a live count.
- Per-LoRA render counts — every adapter shows how many renders used it; click through to the pre-filtered gallery.
- Batch upscale — multi-select ×2/×4 runs through the queue with per-task error isolation.
- “Iteration N / M” labels when a prompt has several seed variants.
- Queue: move-to-top, plus a live “step X / N” and ETA on the running task.
- Grouped sidebar with a live activity dot on Queue while a run is going.
- Large ‹ › arrows in the gallery detail view.
- Health chip — a red status-bar warning when swap or free disk reach risky levels, with recovery advice; disappears when both recover.
- Repair Library — re-syncs the gallery with the Images folder: recipe-bearing PNGs are re-adopted with full metadata, dead records pruned, unreadable strays quarantined — files are never deleted.
- “Remove all data” got stricter — it now also clears settings, the Keychain token and this app's entries in the shared HuggingFace cache, and reports anything it couldn't delete.
- Engine fix (vendored patch): renders no longer collapse into repeated bands when a size transposes the previous one (portrait right after landscape) — a RoPE-cache collision fixed at the source.
- 150+ fixes and hardening changes from two full adversarial audits — download auth hygiene, queue-handoff races, training-state edge cases, atomic writes, honest error surfacing.
Models — two quality tiers (Klein 4B / 9B) plus the base weights LoRA training runs on, each component downloadable on its own.
Shared components — the VAE and the Describe/Auto-caption vision model (4-bit or 8-bit), with an explicit active switch.
Gallery — every render kept with its full recipe; search by prompt, LoRA or seed, filter by model or quality.
Every image keeps its full recipe inside the PNG. Open any result to copy the prompt, upscale it, or “Remix” the exact settings — seed, size, steps, LoRA and all.
System — live CPU / GPU / RAM / MLX / disk, one-click maintenance, and a genuinely complete “Remove all data”.
Everything runs locally. The app keeps no analytics and writes nothing to the macOS unified log.
Your full-resolution renders (each with its recipe embedded) live in
~/Library/Application Support/Typhoonminigen/Images/; logs, the gallery index and thumbnails sit
alongside them (and under ~/Library/Caches/…, ~/Library/Logs/…) and are erasable from the System
tab. Your Hugging Face token is stored in the macOS Keychain; it is sent only to huggingface.co
to authenticate gated downloads.
The models are several GB and live in your ~/Library, and some weights are cached in the shared
~/.cache/huggingface — dragging the app to the Trash does not remove them (macOS can't clean up
after a trashed app). To reclaim the space, open System → “Remove all data” before deleting the
app: one click wipes the models and encoders, this app's entries in the shared HuggingFace cache
(other tools' caches are untouched), your gallery, LoRAs, training data, presets, caches and logs —
and reports anything it couldn't delete. Then drag the app to the Trash.
App code: MIT. Third-party components and model licenses: THIRD_PARTY_LICENSES.md. Klein 4B is Apache-2.0 (commercial use OK); Klein 9B is non-commercial under the FLUX.2 Klein license — using it is between you and Black Forest Labs.
If this is useful to you, you can buy me a coffee — entirely optional, the app is free.
Built on flux-2-swift-mlx (the FLUX.2 MLX
engine) and mlx-swift. Models by
Black Forest Labs (FLUX.2 Klein) and the Qwen team (text encoders);
upscaling by Real-ESRGAN.







