What If AI Could Plan
an Entire Campaign?

Why Change the Default

In most AI editing tools, an uploaded image sits idle until the user writes a prompt. The user diagnoses the image, picks the right tool, describes what they want. An uploaded image usually sits idle until the user decides what to do next.

Image lands on the canvas — Pixie starts analysing immediately Pixie identifies issues and proposes specific actions unprompted User approves — Pixie executes the edit on the canvas Clean output — no prompt was needed

How Pixie Responds

Pixie runs analysis the moment an image lands on the canvas. It surfaces what it finds — "I see many objects in the background. Clean background? Erase background?" — and the user taps to approve. No prompt needed. The user doesn't have to know what tools are available; Pixie tells them what it can see and offers to act on it.

The Coordination Problem

A social media campaign needs several coordinated assets — posts, stories, reels — all on-brand. Generating each one separately means repeating the brand context in every prompt and manually checking consistency across outputs.

How Pixie Plans It

Pixie interprets a single intent into a multi-asset plan. The user approves the plan, provides brand context once, and Pixie generates all assets together. The canvas shows the full campaign at once — the user can compare outputs, swap individual pieces and refine the set before exporting.

What the Workspace Includes

Users generate images and then need to edit them — background removal, upscaling, watermark removal, adjustments, resize — without switching tools.

The Prompt Bar

The prompt bar has 9 variants across Image, Video, and Pixie modes. Users can generate multiple variations from one prompt. Clicking any output reveals the editing toolbar — prompt editing, background removal, upscaling, watermark removal, adjustments and resize — all within the canvas.

The Limitation

Prompt-based editing applies to the entire image. If a user wants to change a dress colour, remove a single object, or adjust one element — they have to describe enough context for the AI to understand what to leave alone. Other tools solve this with mask layers, feathering controls, and lasso selections. Precision editing shouldn't require a Photoshop mental model.

Brush over the element to define the edit boundary Describe the change — prompt anchored to selection Original stays visible while the AI generates the edit Edited image on canvas — full toolbar available for refinement

The Interaction

Four steps: brush, select, describe, apply. The brush creates a spatial mask directly on the image — no layers panel, no feathering controls, no mode switching. The prompt field appears inline, anchored to the selection itself, not in a separate panel. Multiple selections can be stacked before applying, each with its own instruction. That thing, change it to this.

Three Routes, One Interface

Video generation has three different entry points: text-to-video, image-to-video, and motion control. Each calls a different AI model, but the user shouldn't have to know that.

Implementation Detail

All three video types share one flow with mode switching. V2 added a dedicated Motion Control toggle for granular camera parameters. For engineering, video was split into three separate implementations — each maps to a different AI model endpoint — but the user-facing interface stays unified.

How It Works

Same planning pattern as campaigns: the user provides a script, Pixie proposes a scene-by-scene breakdown, and generates each scene as a video clip on the canvas. Individual scenes can be reordered, regenerated or adjusted independently.

Scale

E-commerce teams apply the same edits — background removal, upscale, style adjustment — across hundreds of product images. Repeating that per image doesn't scale.

What I Left Out

The batch toolbar was designed as a subset of the single-image toolbar — 6 variants instead of 7. The missing tool is prompt-based editing, which requires per-image intent and doesn't scale to batch operations. The Template Builder extends this further: save an action sequence and reuse it across future batches.

Consistency

AI-generated influencers need to look consistent across poses, outfits and settings while maintaining a specific demographic and aesthetic.

Structured Input

Instead of a freeform prompt, Pixie runs a structured interview — 5 questions covering niche, demographic, purpose, aesthetic and characteristics. The five inputs mapped directly to the parameters engineering needed for the persona endpoint.