MFLUX-WEBUI

MLX Media

We are back — now as a local media studio.

MLX Media is a private, Apple-silicon workspace for creating, editing, restoring and organizing visual media. It evolves MFLUX-WebUI into one coherent photo-and-video product while keeping model limits, licenses and runtime state visible instead of pretending every backend works the same way.

What works today

Area	Status	Notes
Local photo API	Available	Async image jobs, job history and SSE progress through the existing API
Photo import	Available	Local inventory with EXIF/GPS discovery, explicit privacy modes and manual location overrides
SeedVR2 restoration batches	Available	Folder planning, safe outputs, progress and original-file preservation
React studio	Available	Premium dark interface for Home, Images, Video, Restore, Time Lens, Library, Models, Activity and Settings
MLX video	One verified mode	Local Wan 2.1 1.3B text-to-video through an isolated, pinned runner

The source checkout and an installed Pinokio application are separate. Updating this repository does not restart or mutate an already-running restoration job.

Product principles

• Local by default. Source media stays on the Mac unless the user explicitly chooses another path.
• Originals are protected. Restoration and edits produce new artifacts instead of overwriting source files.
• Location is opt-in. EXIF/GPS may suggest a location, can be disabled per import, and can be replaced with a manual label. MLX Media never presents inferred history as fact.
• One honest queue. Large photo and future video jobs must be serialized to protect unified memory and predictable performance.
• Capability-driven UI. Models expose their own supported operations, safe defaults, constraints, cache state and license requirements.

Photo capabilities

The backend is built around MFLUX on Apple silicon and includes the established async generation API plus local photo inventory and SeedVR2 batch restoration. Exact request and response formats live in API.md.

Highlights:

• text-to-image and image editing through the configured MFLUX model paths;
• asynchronous jobs with status, history and Server-Sent Events;
• SeedVR2 3B/7B restoration and upscale planning;
• non-destructive batch outputs with metadata/orientation handling;
• local photo inventory with HEIF support and privacy-aware GPS suggestions;
• model-specific settings rather than a single unsafe universal form.

Video: one verified local path

The video workflow uses a pinned, isolated Blaizzy/mlx-video subprocess. Only the exact Wan 2.1 1.3B text-to-video profile that passed a local Apple-silicon smoke test is exposed. Its dependencies do not enter the MFLUX photo environment.

Therefore:

• the server owns the capability limits, runtime state and safe model paths;
• jobs share the existing serialized media queue and expose real progress/cancellation;
• completed MP4s and provenance remain local and are validated before publication;
• unsupported LTX, Wan 2.2, conditioning, audio and arbitrary-size modes stay hidden;
• setup is explicit and reproducible; the frontend never silently downloads a model.

Read the full MLX video integration audit.

AI assistance: capability-first

The repository contains several separate local-assistance building blocks, but they are not interchangeable:

• the current Nativ integration only checks a configurable local server’s health and detected model IDs through GET /api/v1/providers/nativ;
• prompt refinement exists in legacy backend code through Ollama or local MLX paths, but is not wired into the React studio yet;
• local MLX-VLM captioning code exists, but is not exposed as a tested automatic-tagging product flow.

For that reason the interface presents Nativ as status integration only and keeps prompt refinement labelled as unconnected. It does not claim a Nativ assistant, image understanding, reliable automatic tags or structured creative assistance.

Run from source

Backend API

Requirements: macOS on Apple silicon, Python 3.11+ and enough local memory/disk for the chosen model.

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt
python api_main.py

The local API binds to 127.0.0.1:7861 by default. Use webui.py as the Pinokio- compatible entry point; it currently starts the same API server.

To allow local photo inventory and restoration planning, configure a dedicated import root before starting the API:

export MFLUX_PHOTO_IMPORT_ROOT="/absolute/path/to/photos"
python api_main.py

The photo endpoints reject non-loopback access and paths outside that root.

React studio

cd frontend
npm install
npm run dev

The React app is capability-driven. Restore and the verified video mode use the local API; unfinished surfaces such as Time Lens stay visibly unavailable.

The interface ships in English, Dutch, Spanish, German and French. English is the default until a person chooses another language in Settings; that choice is stored locally in the browser and restored on the next visit.

Production build:

cd frontend
npm run build

Translation workflow

All interface copy lives in frontend/src/i18n.tsx as semantic message keys. The English dictionary defines the complete key set, and TypeScript requires every other dictionary to provide the same keys. UI components call t("section.message") instead of embedding display copy directly.

To add another language:

1. add its language code and endonym to languageOptions;
2. add a complete dictionary for that code (do not ship machine-translation placeholders);
3. run npm run build in frontend to verify key parity and types;
4. test every route at desktop and mobile widths, including focus labels, dynamic video states and long button text;
5. change languages in Settings, reload, and confirm the saved choice and document language are both restored.

Keep model names, file formats and server-provided technical identifiers unchanged. Translate their surrounding labels and state descriptions instead.

Main API surfaces

• GET /api/v1/health — server and capability health
• POST /api/v1/generate — asynchronous generation and photo-batch submission
• GET /api/v1/jobs — job inventory
• GET /api/v1/jobs/{id} — job state
• GET /api/v1/jobs/{id}/stream — Server-Sent Event progress
• DELETE /api/v1/jobs/{id} — cancellation request
• GET /api/v1/photo-imports/config — local import policy
• POST /api/v1/photo-imports/inventory — metadata/privacy-aware inventory
• POST /api/v1/photo-batches/plan — safe SeedVR2 batch plan
• GET /api/v1/video/capabilities — exact supported local video mode and limits
• GET /api/v1/video/status — isolated runtime, model and queue readiness
• GET /api/v1/video/artifacts/{id}/{file} — validated local MP4 and provenance

Development checks

python -m unittest discover -s tests -v
cd frontend && npm run build

Relaunch notes

• New product name and navigation: MLX Media.
• A single premium dark visual system across desktop and mobile.
• General-purpose examples with no destination-specific hardcoding.
• Privacy-first photo import and Time Lens language.
• First-class Restore workflow for SeedVR2.
• New Video and cross-media Models surfaces with honest readiness gates.
• Shared Activity and Settings concepts for future photo/video resource scheduling.

Licensing and attribution

This repository’s code license does not replace the terms attached to model weights. Review the repository license plus every selected model’s license before downloading or using it. In particular, LTX-2 uses its own community license, while official Wan 2.1 and Wan 2.2 weights advertise Apache-2.0 terms.

Core upstream projects include:

• MFLUX
• MLX
• Blaizzy/mlx-video — pinned, isolated video engine for the verified Wan profile
• Pinokio

Safety note

Large local models can consume substantial unified memory and disk. Start with conservative presets, keep the application local-only, preserve originals, and verify the output of any generative or restoration model before relying on it.