Qwen Prompt Assist
Install the public Qwen 3.8 model and its private runtime in one guided action, then refine model-aware prompts locally.
Build reusable Concepts. Generate images and video. Restore, upscale, compare, and preserve every creative decision in one private production console.
THE REAL STUDIO // V0.9.9
Generate, direct motion, review outputs, and preserve the creative trail without leaving the project.
Install the public Qwen 3.8 model and its private runtime in one guided action, then refine model-aware prompts locally.
Text, image, first + last frame, and performance-audio workflows on Apple Silicon with native MLX.
Build reusable characters and styles, then attach compatible LoRAs to the Concept that owns them.
Drop images from Finder straight into start and end frame slots, then direct the transition between them.
Run the dedicated LTX 2.5 upscale path with visible queue progress and preserved source audio.
Keep settings, sources, render metadata, and every derived version attached to the finished asset.
FIELD TEST // 03:33
A complete ALi Imagine production built from numbered first-and-last keyframes, reusable character Concepts, local LTX motion, immutable outputs, and a final editorial pass.
Watch on YouTubeTHE INSTRUMENT
ALi Imagine surrounds the work with only what matters: sources, signal, engine state, outputs, and complete lineage.
ONE CREATIVE TRAIL
No disconnected model buttons. Four clear surfaces, one project, and a recoverable creative trail.
Collect references, character sheets, guidance, datasets, and certified adapters as reusable creative knowledge.
Generate images and video locally or through xAI without changing the project’s core workflow.
Compare, favorite, derive, reveal, export, and revisit every immutable version from a full-screen media surface.
Install and verify local engines, models, credentials, hardware tests, and capability certification explicitly.
CONCEPTS // REUSABLE IDENTITY
References, character sheets, guidance, and certified adapters become project knowledge. The engine may change. The creative identity remains traceable.
Every frame keeps the path back to what made it.
FIRST + LAST // LTX-2.5 MLX
Build controlled transitions between numbered keyframes. LTX receives the exact opening and closing composition while the project preserves both sources, the generated clip, and its complete recipe.
PROVIDERS ARE PLUGINS
The project and its assets do not belong to an API. Engines execute recipes; they do not own the system.
Fast Apple Silicon image generation, image transformation, and Concepts guidance.
METAL · BF16 · OFFLINEChoose Distilled for speed or Base at 50 steps with CFG 3.5. Draft 512 through native 4K, generative 2× upscale, ten references and layout boxes.
MLX / METAL · 5.6B · 16 STEPSLocal prompt refinement and source-frame analysis with a public, resumable model install that needs no Hugging Face token.
MLX SERVE · 4-BIT · OFFLINEText-to-video, image-to-video, first-and-last-frame control, native audio, and adjustable diffusion steps.
METAL · BF16 · ADJUSTABLE STEPSPrepared FL2VA and Ref2VA integration. Local H3 execution is disabled in public builds pending authorization; private tests do not establish public availability.
PREPARED · AUTHORIZATION-GATEDCreative 2× detail for generated and imported videos. Keeps aspect, timing, frame count, source audio, and immutable originals.
2 STEPS DEFAULT · ADJUSTABLE 2–8 · METALRemote image generation, editing, video generation, video editing, and extension through one adapter.
KEYCHAIN · ASYNC · LOCAL DOWNLOADASSET-FIRST
Every output retains its source, signal, seed, engine, model, adapter, settings, checksum, timings, and parent version. Originals are never overwritten. Every operation derives.



CODEX + DAVINCI RESOLVE / MACOS
Capture a frame, prepare an ALi video shot, review the result, and place the take back on your Resolve timeline. The ALi Resolve skill gives Codex the workflow. ALi Imagine MCP adds engine status, supported settings, source imports and persistent generation jobs through an authenticated local connection.
codex mcp add ali-imagine -- python3 "$HOME/Library/Application Support/ALi-Imagine/mcp/server.py"
Launch ALi Imagine once to install the client, then run this command and reopen your Codex chat. Python 3 is required. The client reads a user-only local token; never paste API keys into MCP configuration. Check ali_status and engine availability before submitting. Source imports, first/last-frame and lip-sync workflows retain request and job identity; results require review before timeline placement.
ali-resolve folder in Downloads.Install the ALi Resolve skill from ~/Downloads/ali-resolve into ~/.agents/skills/ali-resolve. Read its README.md, back up any existing version, and check the required software. Do not generate media or change my timeline.
Use $ali-resolve to check my ALi and Resolve connections. Tell me the active project and timeline. Do not generate anything or change my timeline.
Requires the ALi Resolve plugin, Python 3, and Resolve external scripting access. Generation needs a configured, ready runtime. Public H3 execution remains authorization-gated; experimental settings are not a promise of model availability. This ZIP contains the Codex skill, not the apps, plugin, or models. Start with the connection check; older ALi versions may need an update.
No Sirius or ChatGPT manifest.json is needed for this workflow. Current helper: macOS, whole-number timeline frame rates. Download checksum.
ALi IMAGINE v0.9.43
Grok now offers Image 2.0 quality, Full/Lite video choices, fourteen Full video references, timed middle frames and up to three preset voices. Character Guidance can fill empty continuity fields from references while preserving your edits. Use original reference images directly as Qwen Looks. Choose and order up to ten references in the main Concepts image grid, with automatic saving and covers made from your selected images. Qwen can borrow subjects and composition as well as style; choose references to suit your scene.
Includes Qwen Turbo, Full and Unlocked, QWEN 2×, Refine, raw prompting, FLUX 3, Explore and the earlier video workflows. Existing projects, models, MCP, Resolve and paired-Mac workflows remain available. Qwen uses the noncommercial Qwen Research License. FLUX models require separate Hugging Face access and setup. H3 execution remains authorization-gated. Qwen 2× now preserves inherited Look continuity metadata. The library avoids duplicate style cards, and stored LoRAs remain available in Concepts. Explore orbit remains experimental: reliable full 360° coverage and frozen geometry are unverified. See the release notes for limits.
NOTARIZED RELEASE Developer ID signed and Apple notarized for direct macOS download. No Control-click exception is required.