Connect generation nodes on the Infiknit AI Canvas
Connect source, reference, image, and video nodes in Infiknit without hiding dependencies or overcomplicating the canvas.
Mira ParkVisual Systems EditorConnect generation nodes only when one node supplies a real input or constraint to another. On an Infiknit canvas, a useful connection tells the next reviewer what role an upstream item plays: prompt, image, video, style, character, product, background or camera angle. A useful graph is not the one with the most lines; it is the one where each downstream generation can be understood without guessing. If you cannot explain an edge in one sentence, leave it disconnected until the dependency is clear.
Begin with the destination. If you are making an image, decide whether it needs a prompt, an approved product reference, a character reference, a style direction or some combination that the selected model actually supports. If you are making a video, decide whether it starts from an image, a video source, a motion instruction or a selected frame. Then connect only those sources. The connection is a declared dependency, not an ornament.
Read a connection as a sentence
Before drawing an edge, say it aloud: “This approved product reference informs this image candidate.” “This selected still becomes the source for this motion test.” “This style reference informs this alternate scene.” If the sentence is vague—“these are related”—do not connect the nodes yet.
The product’s canvas model uses typed nodes. Its documented connection roles include prompt, image, video, style, character, product, background and camera angle. A role matters because it preserves intent. A Product reference is not just another picture; it is a source whose recognisable physical qualities should be protected. A Background reference expresses a different instruction. Treating both as anonymous image inputs makes later troubleshooting much harder.
Set up an image connection deliberately
For a product still, start with an approved source Image node. Label it with the product version and the reason it is safe to use. If the item needs to recur across projects, save or attach it as a Product reference rather than leaving it as an unexplained upload. Add an Image generation node titled for the intended output, such as vertical countertop hero—warm morning light.
Add the minimum inputs needed for this particular decision:
- A Text node or prompt with the scene and placement.
- A Product reference when recognisable packaging or geometry matters.
- A Style reference when visual language must remain close to an approved direction.
- A Background reference when the scene itself, rather than merely its color, matters.
- A Character reference only when an actual character must recur.
Check the selected image model before queueing. Available image and video models are filtered by enabled model IDs and validated providers, and capability data includes matters such as accepted modes, resolutions, aspect ratios and reference support. A connection can be conceptually sensible yet unavailable to the chosen model. Do not pretend a model will honour an input merely because the graph is visually connected.
Connect an image to a video without losing context
An approved still can be the starting point for a video test, but it should not become an unquestioned final frame. Add a Video node that says what is changing: camera move, subject action, duration and delivery shape. Keep the accepted image connected to it, and retain the product or character source nearby if identity review will matter.
This layout creates a useful chain: approved source → selected still → motion attempt. If the video changes packaging, facial identity or text, trace back through the chain. If the image was sound, revise the video prompt or model choice. If the image already carried a problem, repair the earlier image branch rather than layering another instruction on the video.
The application can expose video duration, resolution, aspect ratio, reference-image support and other provider-specific capability fields. Review them in the live controls. A video node’s title should never imply that a duration or mode is available until you have checked the selected provider.
Keep branching legible
When you need variations, branch from a stable source and name the one intentional difference. Approved bottle → image: clean studio and Approved bottle → image: rain-lit street communicate two deliberate experiments. A chain that repeatedly connects output to output without labels quickly becomes impossible to audit.
Do not connect every candidate to the final video node. First review candidates, mark the accepted still, and connect that one. This prevents accidental use of an unapproved experiment and keeps the true decision path visible. If two candidates are both valid alternatives, create two video branches and compare the results honestly.
For a broader method, use the AI creative canvas guide. For image-specific branching, see build node-based AI image generation graphs, and for motion work see build node-based AI video generation graphs.
Connection checks before you queue
Run this short review before submitting generation work:
| Check | Why it matters |
|---|---|
| Source is approved and current | Avoids reusing an obsolete product, character or brand asset. |
| Edge role is clear | Lets another creator understand why the input is present. |
| Target is named for its output | Makes results searchable after several iterations. |
| Model accepts the intended inputs | Prevents a visually plausible but unsupported configuration. |
| One variable is being tested | Makes the output useful evidence, not random variation. |
| Queue target is explicit | Avoids running unrelated work and spending unintentionally. |
Queueing does not mean a result is correct. It asks the application to execute an eligible target. Review queue state, provider errors and returned media. A successful node can still fail the creative brief.
What connection validation should prevent
A visual editor should stop invalid or unsafe graph changes before they corrupt the project. Infiknit’s agent and canvas-operation design validates node types, node-specific settings, existing references and connections. The model does not mutate React state directly; the frontend applies validated canvas transactions.
In practical terms, validation should stop attempts to attach an unavailable asset, make a connection to the wrong type of target, or rely on model controls that are not currently supported. Errors are not a reason to add more unrelated references. Read the missing requirement, change the appropriate input or select a suitable model, then retry the smallest affected branch.
Cycles are also a sign to pause. A node should not depend on an output that ultimately depends on itself. If a project looks circular, split it into a source branch and an alternate refinement branch. The goal is to preserve provenance, not construct a puzzle.
A compact example
For a creator-style product clip, use a Product reference, one Image candidate and one Video candidate. The product asset connects to the Image node as the stable object. The image node is reviewed for packaging and composition. Only after it passes does it connect to the Video node. The video prompt asks for a specific move, not a new product design. If the video fails, the still remains intact as a known-good source.
This gives you a clean diagnosis: product issue, image issue or motion issue. It is far more useful than re-running a long prompt that blends every stage.
Limits to state clearly
Connections preserve context; they do not guarantee consistency. Models may interpret references differently, provider capability surfaces can change, and an accepted image may not make a convincing video. The creator remains responsible for rights, claims, visual approval and publication. Use the canvas to show evidence of those decisions, not to hide them.
For agent-led setup, how an AI creative canvas agent works explains the distinction between a visible validated canvas action and an autonomous creative decision. When the same set of connections has proven useful, document it through a reusable creative Blueprint rather than assuming an accidental graph is a template.
FAQ
Should every reference connect directly to every generation node?
No. Connect a reference only when that generation needs it. Overconnecting can make a graph difficult to read and can obscure which input influenced a result.
Can I connect a candidate before reviewing it?
You can, but label it as a candidate. For work that will be reused or sent downstream, connect an explicitly approved output after reviewing it at a useful size.
Does a connected graph run automatically?
Not necessarily. A connected graph describes dependencies; it does not mean every downstream target is automatically queued. Select an eligible target, check its current controls and dependency state, and confirm the returned media before treating the next node as ready. This distinction matters when an image branch is approved but the video provider does not accept that image mode.