What is Canvas AI?

Canvas AI is Archinode's classic canvas: the base image, the instructions, the render engine and the result, already connected, to render a photo without building a flow. It opens from the “Apps” panel on the Dashboard or with “New Project” in Projects.

You upload the base (up to 20 MB; it's reduced to 2048 px on its own), write the instructions — with preset styles, the improve-prompt button (on paid plans) and the “Advanced settings” — and in the engine you pick the model, the resolution (2K from Basic and 4K from Standard), the aspect ratio and 1 to 4 variants, which are generated at the same time; on Archinode AI and NanoBanana PRO the variants after the first cost less (10 and 25 tokens: four come to 50 and 120). When you press “Generate” the button turns red and counts down a few seconds before anything is sent: press it again in that time to cancel at no cost. With the “+” on the engine's “References” pill you add images — and choose each one's type: material, furniture, element, scene, style or winning render —, and a result connected to that pill picks its type the same way; you can also bring pieces from the furniture catalog, and with “Mark reference areas” you mark where each one goes.

The instructions node has two tabs: “Prompt”, where you write, and “Assistant”, the same one as in Nodes, which reads the engine's base and references and writes the prompt for you (in English; it spends no tokens and is for paid plans); with a reference connected you choose which image the camera comes from. The active tab always sits against the node's left edge, and picking the other swaps them. The engine uses only the tab that is active: with the Assistant, its prompt goes to the render as written, without the “Master prompt” or the “Advanced settings”, and if it hasn't written anything yet, “Generate render” does it before rendering.

The engine has two tabs too: “Render” and “Video”. On “Video” the same inputs make a clip instead of an image: the base is the first frame, the text of the active tab describes the movement and, of the images you connect to the third pill — called “Last frame” there — you pick one for the clip to end on (only on the models that use one). You pick the model, the duration, the resolution and, if the model offers them, sound and the advanced settings; the price shows before you generate. Video starts at Standard, and each tab keeps its own settings. A clip takes a few minutes — you can keep working or cancel it: the tokens are returned if it hadn't started yet — and it shows up in the connected result, which plays it and keeps it in its history next to the renders.

Each node carries its name on the tab at the top and, on its first row, its tools: a reference's type, the instructions' “Master prompt” (or the model, on the Assistant tab) or the engine's aspect ratio. The arrow at the engine's top right minimizes it: it folds up and leaves only the model, the resolution, the variants and a round button to generate, and the same arrow opens it again. With the “+” on the bottom bar you add any other node and connect it yourself, and if you drag a connection out of a node and drop it on an empty spot of the canvas, the same menu opens with only the nodes that fit there: the one you pick appears where you dropped it, already connected; selecting one shows duplicate and delete, and lights up its connections while the rest stay faint. Double-click an image to view it large with its details and, with “Edit”, in the same image editor as Nodes (edit with a prompt, enhance, retouch, add people, crop or expand, upscale, change the angle or the light, adjust): what you save goes into the result's history, or replaces the image of a base or a reference, and Ctrl+Z undoes it. Double-click a clip to open Nodes' video panel, to enhance it, add slow motion, a voice or music, modify it, extend it, trim it or join clips; each new version is added to the result's history. Drag a node's corner to stretch it without losing the image's proportions.

In the top bar, the logo opens the menu (Dashboard, new project, AI Chat, styles, community, history, save and support), and next to it is the project's name, which you rename right there. On the right, the ring around your photo shows how many of your plan's tokens you have left (press it to see the number), the bell brings Archinode's updates, “Chat” opens the support chat and “Share” publishes the whole canvas to Community as a flow, just like in Nodes.

The result keeps a history of its last 20 renders (each variant gets its own node): hover it to see them as thumbnails at the top, and click any one to go back to it. It comes with a before/after comparison —against its base or, when its engine has references, the one you pick—, download as PNG, JPG, WebP or PDF, publish to Community and “Use as base” (a clip downloads as MP4 and has neither the comparison nor “Use as base”), which asks how to go on: as a base, it places a new engine to its right with that render as its base, plus an instructions node, to keep iterating; as a reference, the new engine gets the render on its references pill and comes with an empty base image, so you upload another view of the same project and the render guides it. The project saves when each render finishes or with “Save”.