A native macOS Swift app for FLUX, Krea 2, Z-Image, and (prototype) Ideogram 4 image generation powered by mflux and Apple MLX. Queue jobs, watch generations unfold step-by-step, cull and compare results in a Lightroom-style gallery, and even write prompts with a local LLM — all without a CLI, and NOT yet another Electron container "app".
Requires macOS Tahoe 26.0+ and Apple Silicon M-series. mflux is installed automatically on first launch.
| Main window | Live preview |
|---|---|
 |
 |
| Gallery | Settings |
|---|---|
 |
 |
- Text-to-image and image-to-image generation via FLUX.2 Klein, Krea 2 Turbo, and Z-Image models
- Krea 2 Turbo — fast photographic model family, with img2img support
- Z-Image & Z-Image Turbo — Tongyi's 6B single-stream model; the distilled Turbo (9-step, guidance-free) and the base variant (classifier-free guidance + negative prompt), both with img2img and LoRA
- Ideogram 4 (prototype) — structured-caption generation with a regional bounding-box layout editor and Gemma-assisted caption authoring (see Ideogram 4 below)
- Warm-model engine — keeps the model loaded between generations, eliminating reload time on consecutive runs
- Scenario generator — turn a rough outline into finished prompts with a local LLM (no API key, no network round-trip)
- Step-by-step live preview — watch the image denoise in real time
- Persistent job queue — queue multiple jobs (processed FIFO), they survive app restarts
- Lightroom-style gallery — cull with pick/reject flags, filter by metadata or model family, and compare results side-by-side; thumbnail cache and metadata sidecars keep scans fast
- Prompt history & notepad — searchable, pinnable prompt history plus a markdown notepad for reusable notes
- Wildcards & batches — independent wildcard sampling, spread-across-batch seeds, and img2img prompt adoption
- LoRA support — add any number of LoRA adapters with per-adapter strength sliders and reordering
- Batch generation — run 1, 3, 5 or a custom count with auto-incrementing seeds, plus a learned generation-time estimate
- Shortcut Keys -
Command + Enterfor one-shot generation,Option + Command + Enterfor batch generation - Model defaults — save per-model presets (steps, guidance, LoRAs, dimensions)
- Prompt templates — save and reuse favorite prompt fragments like lighting, camera style
- Metadata tools — token counter with a 512 soft-cap for FLUX.2, and one-click stripping of embedded metadata from exported images
- Low-RAM mode — streams transformer blocks to cut peak Metal memory ~75%
- HuggingFace integration — enter your HF token once (stored in Keychain); gated models download automatically
| Variant | Steps | Notes |
|---|---|---|
| FLUX.2 Klein 4B (distilled) | 4 | Fastest |
| FLUX.2 Klein 9B (distilled) | 4 | Best quality/speed trade-off |
| FLUX.2 Klein 4B (base) | ~50 | Full diffusion |
| FLUX.2 Klein 9B (base) | ~50 | Full diffusion |
| Krea 2 Turbo | ~few | Fast photographic model; text-to-image and img2img |
| Z-Image Turbo | 9 | Tongyi's distilled 6B single-stream model; guidance-free, text-to-image and img2img; BF16/Q8/Q4 (Q4 loads a pre-quantized MLX repo) |
| Z-Image (base) | ~50 | Base Z-Image with classifier-free guidance and a negative prompt |
| Ideogram 4 (prototype) | preset | Structured-caption model; FP8/Q8/Q4 precision selector (gated repo — accept terms on HuggingFace). Requires unreleased mflux support — see below |
| Custom | any | Any HuggingFace repo ID or local path, loaded as any model above — pick the architecture it matches under "Loads as" |
Prototype: Ideogram 4 support is fully built out in the app, but generation depends on mflux CLI support that has not landed in a published mflux release yet (see the note below). Until then, treat it as a preview of the editor and workflow rather than a working generation path.
Ideogram 4 is a structured-caption model: instead of a single prompt string it takes a JSON caption describing a high-level scene, an optional style block, and a compositional breakdown of regional elements (each with an optional bounding box and color palette).
- Caption editor — author the structured caption in a sectioned form, or let
Gemma turn a plain description into a full caption (
mlx_lmruns locally viauv; no API key). - Bounding-box layout editor — drag, resize, and color regional elements on a canvas overlaid on the output aspect ratio.
- Color palettes — per-element and per-style palettes with editable hex entry, so an exact color can be reused across elements.
- Precision — FP8 / Q8 / Q4 selector. Q8/Q4 load pre-quantized MLX weights
directly; FP8 quantizes once via
mflux-save.
mflux support: Ideogram 4 generation drives the
mflux-generate-ideogram4CLI. Any model whose CLI is missing from your mflux install is disabled in the model picker with a note saying so, rather than failing when you hit Generate — which family CLIs exist varies by mflux version. The model is gated on HuggingFace — accept the terms on the model page before first download.
| Requirement | Version |
|---|---|
| macOS | Tahoe 26.0+ |
| Apple Silicon | M1 or later |
mflux and uv are installed automatically on first launch if not already present. A progress banner appears at the top of the window while the install runs — no manual setup required.
Download the latest MLXBits_Image_Studio_<version>.dmg from Releases, open it, and drag MLXBits Image Studio to /Applications.
The app is notarized/signed by Apple, so it should launch without any Gatekeeper prompt.
Like the app? Support me by buying me a coffee. :) Your support helps keep my apps and other content free.
Requirements: Xcode 26+, XcodeGen, SwiftLint, SwiftFormat.
brew install xcodegen swiftlint swiftformat
git clone https://github.com/MLXBits/image-studio
cd mlxbits-image-studio
# Generate the Xcode project
xcodegen generate
# Open and build in Xcode
open "MLXBits Image Studio.xcodeproj"Signing: Set
DEVELOPMENT_TEAMinproject.ymlto your 10-character Apple Developer Team ID before building a signed Release. Debug builds are unsigned and work without a Team ID.
Releases are built, signed, notarized, and published automatically by GitHub Actions
(.github/workflows/release.yml) whenever a vX.Y.Z tag is
pushed. The tag name is the source of truth for the release version.
git tag v0.6.3
git push all v0.6.3 # → triggers the release workflowThe workflow derives the version from the tag, builds the Release archive with Developer ID
signing, creates a notarized + stapled DMG, and publishes a GitHub release with generated notes.
The tag is the sole source of truth for the version. (Optionally bump MARKETING_VERSION in
project.yml too, so local Debug builds report the new number — CI overrides it either way.)
One-time setup — add these repository secrets (Settings → Secrets and variables → Actions):
| Secret | Value |
|---|---|
DEVELOPER_ID_CERT_P12_BASE64 |
base64 < DeveloperID.p12 — the "Developer ID Application" cert exported with private key |
DEVELOPER_ID_CERT_PASSWORD |
the .p12 export password |
DEVELOPMENT_TEAM |
your 10-char Apple Team ID |
APPLE_ID |
Apple ID email for notarization |
APPLE_APP_SPECIFIC_PASSWORD |
app-specific password from appleid.apple.com |
App/ App entry point and root layout
Models/ Job models, model catalog, LoRA entries, prompt history/templates
Runner/ Generic JobRunner<Spec> engine, per-family runner specs, warm-driver controller
Stores/ @Observable state — AppSettings, per-family job stores, GalleryStore, TimingStore
Views/ SwiftUI views (ParamsPanel, PreviewPane, Gallery, Queue, Settings, Ideogram4, Krea2, ZImage)
Utilities/ KeychainHelper, MetadataSidecar, progress parser, caption/scenario LLMs, installers
Tests/ Swift Testing unit tests
Resources/ Info.plist, entitlements, bundled Python drivers
project.yml XcodeGen source of truth (never edit .xcodeproj directly)
For naming conventions, the per-family file layout, and a step-by-step recipe for adding a model family, see AGENTS.md.
AGENTS.md is the canonical orientation file, read by Cursor,
Codex, and others. Claude Code looks for CLAUDE.md, so after cloning, create
a local symlink:
ln -s AGENTS.md CLAUDE.mdThe symlink is gitignored on purpose — one tracked file, no duplicated guidance to keep in sync, and no tool-specific artifact in the repo.