The software is free and open source, and it runs on your own computer. You pay for the hardware, and for Partner Nodes if you call closed models such as Google’s or OpenAI’s, which draw on prepaid Comfy credits. Comfy Cloud, which runs it on rented GPUs, starts at $20 a month.

Can people who do not build canvases use one?

Yes, once the builder publishes it as a simple form. Figma Weave calls these Tools, Flora has App Mode, Krea has a Node App Builder and Runway publishes workflows as apps. The person running it fills in the inputs and never sees the wiring.

Do I need a canvas to use several AI models?

You can call several models without one, through each maker’s own app or API, or through OpenRouter, a service that gives one account and one bill for many models. A canvas adds the wiring between the models.

What are node-based AI workflows, and do you need one?

Checked September 24, 20266 min read

Short answer

A node-based AI workflow is a canvas of boxes wired together: inputs such as product photos and a brief, AI steps such as a text model that writes prompts and an image model that renders them, and outputs. Agencies build them to run the same pipeline on every product and to mix models from different companies. You need one if you run many similar jobs and someone on the team can build and look after it.

What a node canvas is

A node canvas is a page of boxes, called nodes, joined by wires. Each box does one job: it holds an input, runs an AI model, or edits and exports a result. The output of one box flows along a wire into the next. Once the chain is built, you swap the inputs and run it again.

Node type	What it does	Examples from the vendors’ docs
Input	Brings in an image, text, video or file	Text prompt and Figma frame nodes in Figma Weave; Text, Image and Document nodes in Flora
Language model	Reads the inputs and writes text, such as prompts or image descriptions	LLM, Any LLM and Image Describer in Figma Weave; LLM nodes in Runway and Krea
Image or video model	Makes pictures or clips from text and reference images	Nodes for Nano Banana, GPT Image, FLUX, Kling, Veo and others
Edit	Changes a result: layers, masks, relighting, upscaling	Layers, masks, relight and upscale tools in Figma Weave; upscaler nodes in Krea
Batch and routing	Runs many inputs through the same steps, or sends one input to many boxes	Batch and Router in Flora; List, Array and Text iterator in Figma Weave
Output	Collects results, exports them, or turns the canvas into a simple form	Output in Figma Weave, which unlocks its Tools; Export in Flora, to downloads or Google Drive

Why agencies build them

The same steps run the same way on every product. A new product means a new input and the same process.
Many models sit on one canvas. Flora lists more than 60 models on its pricing page, Figma Weave names Google, OpenAI, ByteDance, Black Forest Labs, Kling and Runway among others, and Comfy’s Partner Nodes reach 38 providers.
Instructions can be locked. A text model with fixed instructions writes every prompt to the same rules, so the set stays consistent.
One branch can run again alone. When one shot fails, you rerun that branch instead of the whole chain.
A finished canvas can go to people who do not build. Figma Weave calls these Tools, Flora has App Mode for its Techniques, Krea has a Node App Builder, and Runway publishes workflows as apps.

The shape of a campaign canvas

Our agency built four campaigns by hand in Figma Weave in 2026, and every canvas had the same three zones.

Inputs: the product photos, reference images and the brief, each in its own box.
A locked brain: a chain of AI roles, each with fixed instructions and its own job. Locking the instructions meant every run started from the same rules.
An output fan-out: about ten shot boxes fed by the brain, each rendering one shot of the set.

Across those four campaigns, 361 renders gave 45 photos an art director kept, about 8 per photo. The canvas made every run repeat the same steps, and still the brownie job, with motion, took 12 renders per kept photo and the surf job, with one model in three locations, took 13.6. See how many renders per usable photo.

The main canvases in September 2026

Tool	What it is	Entry price	Worth knowing
Figma Weave	Weavy, bought by Figma in October 2025. A browser canvas for image, video, audio and 3D models, with editing tools.	Free plan with 150 credits a month; Starter $24 a month, or $228 a year	Your own API keys only on Enterprise. Weave tools also run inside Figma Design, in open beta since June 2026.
Flora	A canvas for designers, brand teams and agencies. Its FAUNA agent builds nodes from a stated goal.	Free, up to 17 generations; Starter $18 per seat a month	Usage is "metered at published model rates", and the seat price comes back as usage budget. Pentagram and AKQA are named users.
ComfyUI	An open-source node engine (GPL-3.0) that runs on your own computer. Comfy Cloud runs it on rented GPUs.	Free on your own machine; Comfy Cloud Standard $20 a month	The most flexible and the most technical. Closed models such as Google’s and OpenAI’s run through paid Partner Nodes.
Krea Nodes	Krea’s canvas, alongside its own Krea 2 image models and training on your images.	Nodes come with the Pro plan: $35 a month on the pricing page	Its Node App Builder turns a workflow into a form. Krea’s docs put Pro at about $30, so check the price at checkout.
Magnific Spaces	A canvas from Magnific, the company formerly called Freepik.	Premium $20 a month	Comes with Magnific’s stock library. Free users can make up to 3 Spaces.
Runway Workflows	A node editor inside Runway, whose strength is video.	Standard $15 a month; Workflows are on paid plans	Workflows can be published as apps, and since July 2026 an agent can build one from plain language.

Prices from each vendor’s own pricing page on September 24, 2026, billed monthly unless stated. They change often.

For a closer comparison of the brand-team canvases, see Higgsfield vs Flora vs Figma Weave.

What it costs and what it takes

Cost	What to expect
Seats	About $15 to $60 a month per person on the entry and team plans above; Flora’s Max plan is $200 a seat
Model use	Credits bundled in the plan, or metered usage. Compare a credit’s real price with the model maker’s own: Google charges $0.067 for a standard 1K Nano Banana 2 image through its API.
Tries	About 8 renders per kept photo in our own campaigns, from about 3 to 14 by job
Builder time	Someone builds and tests the canvas, and fixes it when a model changes or is retired
Hardware	Only for ComfyUI on your own machine. Comfy recommends a dedicated NVIDIA or AMD graphics card for its Windows app.

The skills are prompt writing for each model, knowing which model is good at what, wiring logic such as batches and routing, and the patience to test one change at a time. You can avoid credit markups by bringing your own model keys where a tool allows it; Figma Weave allows it only on Enterprise. See what bring your own key means.

When a brief-in tool is enough

A brief-in tool runs a fixed pipeline behind a form. You give it a product photo and brand details, and it plans and renders the set. Google’s Pomelli, launched in October 2025 as a Labs experiment, reads a website into a "Business DNA" and generates campaign creative; Overs plans and renders art-directed photo sets. You give up control of the wiring and get a pipeline that works on day one.

Build a node canvas if	Use a brief-in tool if
You run the same kind of job every week	You run a few campaigns a year
Someone on the team enjoys building and fixing pipelines	Nobody has time to maintain one
You need video, 3D or models a tool does not offer	Still photos are the job
You want to own every step and every prompt	You want the steps done in a tested order
Other software needs to drive it through an API	Exported files are enough

The wider field, by job, is in the best AI product photography tools.

Questions people also ask

Is ComfyUI free?: The software is free and open source, and it runs on your own computer. You pay for the hardware, and for Partner Nodes if you call closed models such as Google’s or OpenAI’s, which draw on prepaid Comfy credits. Comfy Cloud, which runs it on rented GPUs, starts at $20 a month.
Can people who do not build canvases use one?: Yes, once the builder publishes it as a simple form. Figma Weave calls these Tools, Flora has App Mode, Krea has a Node App Builder and Runway publishes workflows as apps. The person running it fills in the inputs and never sees the wiring.
Do I need a canvas to use several AI models?: You can call several models without one, through each maker’s own app or API, or through OpenRouter, a service that gives one account and one bill for many models. A canvas adds the wiring between the models.

Where Overs fits

Overs is a brief-in tool built on the process behind those canvases: its ten steps run in a fixed order and you watch each one instead of wiring them. It has no public API and does not make video, so a team that needs a custom pipeline or motion is better served by a canvas.

Free for 40 photos a month. The AI that makes the photos is billed separately, on your own key, with no markup from Overs.

Start free How it works

Sources

Checked on September 24, 2026. Prices, specs and rules change; follow the links for the current versions.