Designing with AI

The Canvas AI chat#

The chat panel is the heart of Canvas. Describe what you want and the AI generates new frames or edits the ones you've selected — streaming the result onto the canvas in real time.

• Nothing selected → the AI creates new frames.
• A frame selected → the AI edits that frame in place, preserving its size.
• Elements selected inside a frame → the AI scopes its edits to just those elements.

Massive parallel generation#

Canvas is built for parallel agent generation. Ask for several designs in one prompt and Canvas triggers a separate agent for each — up to 100 running at once — every one building its own frame in parallel:

Make 5 landing pages for a coffee subscription startup

Design 8 mobile app home screens for a meditation app

To keep the results from looking alike, Canvas assigns each frame a unique combination of diversity axes:

• Palette — a distinct color scheme per frame
• Typography — different type pairings and treatments
• Layout — different structural directions (editorial, split-screen, dense list, etc.)
• Navigation — different navigation patterns
• Light/dark — paired with the palette so frames diverge visually

Variations of an existing frame#

Select a frame and ask for variations to re-skin it while keeping its structure:

Make 3 variants of this with different color schemes

Variants reuse the source frame as a reference. If you ask for divergent variations ("each one wildly different"), Canvas drops the structural anchor so each design can drift to its own distinct look.

Guide results with a design system#

To keep generations on-brand, steer them with a design system — your palette, typography, spacing, and component conventions. Provide a design.md file or create a design system from scratch in Omma, and the AI follows it across everything it generates, so even a batch of dozens of frames stays consistent.

This is the cleanest way to get cohesive output at scale: define the rules once, then let parallel generation apply them to every frame.

Choosing a model#

Canvas offers a tiered model picker so you can trade speed for quality on every prompt. The tiers run from the fastest, most efficient options — great for quick drafts and iterating — up to the highest-quality options for polished, detailed designs. Each option is labeled with what it's best for, so pick based on the speed-versus-quality trade-off you want rather than memorizing a specific model.

You can also set the reasoning effort (Low → Max) to control how much the model "thinks" before generating — higher effort produces more considered designs but takes longer.

Scoped vs. full context#

Each prompt can replay either:

• Scoped (default) — only the history for the object you're working on, so edits to one frame don't bleed styling in from unrelated turns. Each object effectively has its own thread.
• Full — the whole conversation, with each turn labeled by the object it targeted.

Importing assets into a design#

Use the "+" button in the chat prompt (or drag and drop) to attach assets so the AI can use real content in your design:

• Images — JPEG, PNG, WebP, HEIC, BMP, and SVG
• Video and audio files
• 3D models — .glb, .gltf, .fbx, .obj, .stl, and more

Assets from your Omma library are passed by reference so the AI can embed them directly (for example, dropping a generated product photo into a landing page). You can also pull in anything you made in Studio.

Background generation#

Multi-frame generations are durable: they keep running on Omma's servers even if you close the editor, and finish in the background. If a stream is cut off mid-design, Canvas keeps a substantial partial result rather than discarding it. Generations consume credits per token; if you run low mid-batch, the frames that completed are kept.