August 12, 2026
What is an AI workflow? The node canvas explained to people who have only known chat
You learned AI creation in a chat window, and you lose your prompts, your settings, and your images with every session. Here is what a node canvas changes: a process laid out visually, permanent, editable piece by piece. A complete explanation, node by node, with an example taken apart cable by cable.
You learned AI in a chat window, and that's normal
Almost everyone discovered AI image creation the same way: a text box, a button, an image that appears. It is simple, immediate, and perfectly fine for getting started. You type "a ginger cat on a Paris rooftop at sunset", you get a ginger cat, you are happy. Conversational chat democratized AI creation; nobody disputes that.
The problem shows up later, usually around the thirtieth generation. You got a gorgeous image ten days ago. The exact prompt? Lost somewhere in a two-mile-long conversation thread. The model you used? No idea. The settings? Vanished. You rebuild everything from memory, and of course the result looks nothing like the original.
This is not an attention problem on your part. It is a structural flaw in the tool. A conversation is linear and ephemeral by nature. But creating an image or a video you genuinely like is a process, with steps, backtracking, and variants. A discussion thread is the worst possible format for representing that.
The three walls of chat
First wall: nothing is reusable. Every conversation starts from zero. Your way of describing light, the prompt structure that works for you, the model that suits your style, all of that lives in your head or in a notes file on the side. The tool itself retains nothing you can build on.
Second wall: no clean variations. You want the same scene in a day version and a night version? In 1:1 for Instagram and 9:16 for stories? In a chat, you resend the prompt with your edits, and you pray the model does not also change the setting, the character, and the composition along the way. It almost always does.
Third wall: no way to fix a single step. If your final video fails because the start image was mediocre, you have to regenerate the image, then dig up the video prompt, then run everything again. Every time, you replay the entire chain by hand. Three edits a day is fine. Thirty is hell.
A workflow is your process drawn on a work surface
A node canvas takes the process that already existed in your head and lays it flat on a large work surface. Each step becomes a box, called a node: here I write my idea, there I generate an image, over there I turn it into a video. Cables connect the boxes to each other and carry the result of one into the input of the next.
The fundamental difference from chat is permanence. Your workflow stays right there, open on the canvas, with all its prompts visible, all its settings displayed, all its images still in place. Come back three weeks later and everything sits exactly where you left it. On Imaginode, saving is automatic; there is not even a Save button.
And because the process is drawn out, it becomes something you can manipulate. You can unplug a cable, swap out a step, duplicate a branch to test a variant. The workflow is not a log of what you did: it is a machine you built, one you can rerun, tune, and improve indefinitely.
The Text node, your externalized brain
The Text node is an LLM sitting on the canvas. On Imaginode, you choose the model running inside it: GPT-5, Claude, Gemini, Kimi K2.5, DeepSeek, and others. You give it an instruction, it produces text. So far, nothing more than a chat. The difference is what you plug into it.
This node accepts text from other nodes as input, but also images to analyze. Concretely: you connect a photo of your product to a Text node with the instruction "describe this product for an advertising image prompt", and it writes the description for you. That description then travels down a cable to an Image node. You have just created a chain where the LLM works for the image generator.
That is a use case you cannot reproduce cleanly in a classic chat, where you would have to copy-paste one tool's answer into another, by hand, on every single iteration. Here, copy-paste no longer exists: the cable takes care of it, on every generation, without ever forgetting or distorting anything.
The Image node, where the pixels appear
The Image node is the generator itself. A prompt field, a model picker, a Generate button, and the image appears in the node, ready to be plugged in elsewhere. On Imaginode, the same node gives access to the whole range: Flux Schnell at 1 credit for roughing out, Seedream 5 Lite at 5 credits when references must be respected, Flux 2 Pro at 7 credits for the final rendering.
The exact cost shows on the Generate button before you click, which rules out unpleasant surprises. One credit is worth about €0.01. Testing four compositions in Flux Schnell therefore costs 4 cents, and you only move to an expensive model once the composition is validated. This discipline of the cheap draft followed by the polished final is probably the most profitable habit you can build.
The prompt field hides a discreet but precious tool: the magic wand. One click, 1 credit, and your rough sentence is rewritten into a rich, structured English prompt, via the Kimi model. Your reference @mentions are preserved along the way. If you do not write English fluently, it is the difference between a decent image and a precise one.
The Video node, the most demanding of the family
The Video node turns an image or a prompt into an animated clip. Its inputs change with the selected model, which makes sense: every video model has its own capabilities. A start image, always just one. An end image, if the model supports it. Reference images to keep a face or a product consistent: 9 at most on Seedance 2.0, more on Seedance 2.5.
Let's be honest, because nobody says it enough: AI video still fails often. Roughly one generation in three or four goes off the rails, a hand warps, an object melts strangely, a camera move ignores your instruction. And an ugly but technically delivered result stays paid for. Only a technical failure on the provider's side triggers an automatic refund, within seconds.
That is precisely why the workflow matters more in video than anywhere else. When a 48-credit clip fails, you do not want to rebuild your prompt from memory. You want to reopen the node, change three words in the motion sentence, and run it again. On a canvas, the rerun costs one click. In a chat, it costs ten minutes of archaeology.
Reference and Style, the memory chat never had
The Reference node solves the oldest problem in AI creation: the character whose face changes with every image. You create a reference with a name, a description, and photos. Then, in any prompt on the canvas, you type @ followed by the name. The mention shows up in green, the description is injected into the prompt, and the photos are attached to the generation automatically.
Your character, your logo, or your product becomes a word. "@Léa sitting in a café, morning light", and Léa keeps her face. Without a reference, let's admit it plainly, she would not: models reinvent a face on every roll. It is the number one limitation of image generation, and references are the only reliable answer.
The Style node plays the same role for art direction. You describe your aesthetic in it and drop in moodboard images. Connected to your Image nodes, it guarantees the entire production shares the same signature. Add the Media node, which accepts your own files by drag and drop, and the canvas finally knows your universe instead of starting from scratch.
The Camera and the comfort nodes
The Camera node deserves an article of its own, but let's summarize. Five illustrated dials: framing, lens in millimeters, aperture, angle, movement. You turn them like on a real camera body, and the node writes the right English terms for you at the head of the prompt, with the movement as a final sentence, right where video models understand them best. A cinematographer's vocabulary, without having to learn it first.
Around these production nodes orbit the comfort nodes: sticky notes for annotating, free text, freehand drawing. It sounds trivial until the day you reopen a month-old workflow and the sticky note that reads "version approved by the client on the 12th" spares you half an hour of doubt.
The canvas itself works for you. Input dots light up when a cable connects, nodes snap into alignment as you move them, and a cable dropped on the body of a node plugs itself into the most likely input. Details, yes. But it is the sum of those details that lets you spend hours on a canvas without getting tired.
A complete example, taken apart cable by cable
Let's take a real case: an animated poster for a bakery's Instagram account. First node, a Reference named @Fournil, with three photos of the storefront and a two-line description. Second node, a Style with four moodboard images, warm tones, film grain. These two nodes generate nothing: they feed the others.
Third node, a Text node running GPT-5: "write an advertising image prompt featuring @Fournil on a winter morning, fogged-up shop window, bread coming out of the oven". Its output travels down a cable to an Image node set to Seedream 5 Lite, 5 credits, chosen because it is the champion of references. The Style node is plugged into the same art direction input.
The validated image moves on to a Video node running Kling 2.5 Turbo, about 26 credits for 5 seconds, with a short motion prompt: condensation sliding down the glass, steam rising from the bread. Total for the chain: five nodes, four cables, about thirty credits per complete version, roughly 30 cents. And above all, the chain still exists tomorrow.
Surgical iteration, the real point of all this
Here is the heart of the matter. In the bakery example, imagine the client finds the image too dark. In a chat, you would start over from the description, hoping to recover everything. On the canvas, you open the Image node, add "golden hour light" to the prompt, and regenerate for 5 credits. The reference, the style, the Video node: nothing else has moved.
Call it surgical iteration: changing a single part of the machine without touching the others. The new visual automatically flows down the cable to the Video node, which waits for just one click to produce the corrected version. A fix that used to take a quarter of an hour of reconstruction now takes forty seconds.
It also changes the way you experiment. Want to compare Kling and Seedance on the same image? Drop a second Video node next to the first, plug in the same input cable, and the two versions generate side by side. Chat forced you to choose before trying. The canvas lets you try before choosing.
Per-node history, your safety net
Every node on Imaginode keeps its last 12 generations. Not a global, jumbled history like a chat thread: a local history, specific to each step. The bakery's Image node keeps its last 12 images, the Video node its last 12 clips, each one viewable right in place.
The most frequent use is dead simple: generation number 7 turned out better than number 11 after all. You recall it with one click, it becomes the node's active output again, and everything plugged in downstream uses it. In a chat, that image would be forty messages deep, and "reactivating" it would not even mean anything.
Add the classic undo-redo for manipulating the canvas itself, and you get an environment where mistakes cost almost nothing. People experiment far more when they know no attempt overwrites the previous one. The 12 generations kept per node have a lot to do with final quality: you dare variants you would never have tried flying blind.
The Photoshop moment: layers for AI
History has already run this experiment. Before 1994, retouching an image was destructive: every brushstroke crushed the pixels underneath, and going back meant starting over. Then Photoshop 3.0 introduced layers. Editing became a stack of independent elements, editable one by one, reorderable at will. No serious professional ever went back to flat retouching.
AI chat is flat retouching: each exchange overwrites the context of the previous one, and the final result is a block you cannot reopen. The node canvas is layers: your creation becomes a structure in which every element stays alive and editable. The comparison is not a marketing line, it is the same conceptual shift, thirty years later.
And as with layers, there is an entry cost. The first minutes on a canvas are less intuitive than a text box, that is a fact. You have to understand the ports, the cables, the direction of the flow. Count on about an hour to get comfortable. Layers took an hour too. We know how that story ended.
Templates, so you never start from a blank canvas
Nobody should build their first workflow node by node, and Imaginode does not ask you to. Ready-made templates cover the most common cases: text to image, image to video, and a complete pipeline that chains prompt writing, visual generation, and animation. You start from a machine that works, and you modify it.
That is the right way to learn. Opening the text-to-image template, you see concretely how a Text node feeds an Image node, where the cables plug in, what a well-structured prompt looks like. You swap in your own subject, you generate, it works. You have just learned the canvas without reading a single page of documentation.
When curiosity strikes, the documentation does exist all the same, with an academy of YouTube video courses built in. But in practice, most users learn by taking the templates apart, the way people once learned HTML by viewing the source code of web pages. Copying a machine that works remains the most effective pedagogy in the world.
The assistant that builds the workflow for you
There is an even more direct shortcut: not building at all. At the bottom right of the canvas, a bubble opens the Imaginode assistant. You write "I want a workflow that turns a product photo into a vertical clip for TikTok", and it builds the entire machine: the right nodes chosen, the prompts written in English, the connections wired. One click, and everything lands on the canvas, laid out cleanly in columns, with no overlap.
The assistant sees your open canvas, which changes everything compared with a generic chatbot. You can ask it why your Video node refuses an input, which model would suit your style better, or how to improve an existing prompt. It answers about your specific case, not a textbook one. Each message costs 1 credit, about a cent.
The irony is delicious: the best argument for the node canvas turns out to be a chat interface. But a chat that produces a permanent structure instead of disposable answers. You talk for a minute, and out comes a reusable machine. It is the exact opposite of classic chat, where you talk for an hour and nothing remains.
When chat is frankly still enough
Let's be honest all the way through: the canvas is not always the right answer. For a single image with no follow-up planned, an article illustration or a quick avatar, a chat or a simple isolated node is plenty. Building a workflow for a one-off generation is rolling out a cement mixer to drive a single nail.
The canvas pays off as soon as there is repetition or a chain. A character to decline across several images, a visual to animate, a series of posts sharing the same art direction, a process you will redo next week. The rule is simple: if you plan to do it again, draw it. If it is a one-shot, generate and move on.
One last piece of candor: on a phone, Imaginode runs entirely in the browser, touch canvas included, and generation continues even with the screen locked. Perfect for rerunning a node or checking a result on the train. But a very large workflow is still more comfortable with a mouse. Touch does the job in a pinch; the 27-inch screen always wins.
Where to start, concretely
The shortest path: open Imaginode in your browser, no installation, everything generates server-side. Load the image-to-video template. Replace the example prompt with an idea of your own, run the magic wand over it for 1 credit, generate the image in Flux Schnell for 1 credit more. You will have understood the principle of cables before spending 5 cents.
In your second session, create your first Reference with three photos, and mention it in a prompt. That is the moment the canvas stops being a gadget and becomes your tool: the machine now knows something chat could never retain. From then on, every workflow you build is an asset that stays.
And if you want to see right away what a complete workflow looks like under real conditions, with the prompts, the camera settings, and the credit bill, we documented the making of a 10-second video ad from end to end in another article on this blog. It is the natural sequel to this one.