🦋 ButterFly — On-Device Image Generator

Your AI art director, running entirely on your Laptop. Describe what you're selling; it interviews you, designs the banner, and hands you a layered PSD. No cloud, no API keys.

What it is

ButterFly turns a few words (and optionally a product photo) into professional marketing & advertising banners, the kind that normally need a designer. Instead of making you write complex image prompts, it acts like a real art director: it interviews you with quick, tappable questions, forms a creative direction, writes the ad copy, generates the image on-device, lets you drag-edit the result, and exports a layered Photoshop file.

Everything runs locally on your machine. Nothing leaves your machine.

🎨 Adobe Creative Track

• A design tool that puts your idea on screen, not the model's. ButterFly doesn't hand you an empty prompt box. It interviews you like an art director: it reacts to each answer, recommends a pick, and builds a creative direction you can watch grow.
• The result stays yours to shape. The headline, label, button and brand are live layers you can drag, resize, recolor and retype on the banner. The layered PSD export (background, detected objects, product and each text block on its own layer) opens straight into Photoshop, so ButterFly fits into a real design workflow instead of ending at a flat image.
• It teaches you to direct AI. Your brief appears as the "starting prompt," and the final page shows the image from your words alone next to the image from the prompt your answers built, with the same seed and size so only the prompt changed. You see exactly what creative direction adds.

🔒 Shock Tracker: Zero Data

• No accounts, no profiles, no tracking. There's no sign-up, no login and no analytics. You open the app and start. Nothing identifies you, because nothing needs to.
• Your data never leaves the laptop. Both AI models run on-device through Apple's MLX. Your product photos, brand name and contact details only go to a server running on the same machine (localhost). There are no cloud APIs, no API keys and no third-party services involved in generating a banner.
• Useful software without a data business model. No subscription, no per-image fee, and no reason to collect anything. After the one-time model download it works fully offline, which proves the point: the whole product works with zero user data leaving the device

Why it's different

• No blank canvas. You never face an empty prompt box. The AI asks the right questions and shows you options.
• No garbled text. The image model renders only the scene; the headline, kicker, CTA, brand and contact are crisp, editable overlays — never "AI text."
• Your actual product. Upload a photo and it's cut out and composited into a generated scene, so the real product (logo and all) appears — not a lookalike.
• Fully local & private. Models run on your Mac via Apple's MLX. No subscriptions, no data sent anywhere.

How it works

1. The interview — a vision-language model (Qwen2.5-VL) runs a short, adaptive interview. It reacts to each answer, has opinions (★ its pick), and asks tailored questions per brief — a café gets "plating / setting / style," an app gets "concept / imagery / style." Almost everything is tappable cards; only a brand name or contact needs typing.
2. The image — your answers are synthesized into a rich prompt and rendered on-device with FLUX.1-schnell. Upload a product photo and rembg cuts it out so it can be composited into the scene.
3. The text — the copy (kicker / headline / CTA / brand / contact) is generated and overlaid, auto-placed and auto-colored to suit each image.
4. Edit live — drag, resize, recolor, and retype any text element directly on the banner; switch layouts; pick fonts.
5. Export — download a layered .psd (background + detected objects + product + one layer per text block) for Photoshop / Affinity / Photopea / GIMP.

Key features

• 🎨 Adaptive art-director interview — reacts to your answers, recommends picks, keyboard shortcuts, a live "creative direction" strip that builds as you go.
• 🖼️ On-device generation — FLUX.1-schnell via MLX; 16:9, square, or story formats.
• 📸 Product compositing — your real product placed into a generated scene (background removal + compositing).
• ✍️ Smart copy — benefit-driven kicker / headline / CTA, with brand & contact pulled through verbatim.
• 🖱️ Interactive canvas editor — drag / resize / recolor / retype text, live.
• 🧩 Layered PSD export — every element on its own editable layer.
• 🔒 100% local — no cloud, no API keys.

Tech stack

Built With

• css
• flux
• html
• node.js
• owen
• python