Build node-based video graphs in Infiknit

A video graph is a shot plan with evidence attached. The useful unit is not “a workflow”; it is a shot brief, an approved source frame, one motion decision, a continuity check, and a deliberate handoff to the next shot. Infiknit keeps those dependencies visible on a canvas so a reviewer can find the source and decision behind a clip.

Short answer: map the sequence shot by shot. Give every shot an action, a start frame, an end condition, and a continuity test. Generate a short candidate, inspect the first, middle, and last frames, and branch when one meaningful variable changes. Trim only from an accepted moment. Hand off a still or clip to the next shot with its settings and review note. A graph exposes dependencies; it cannot guarantee stable motion or identical provider output.

Begin with shot cards

Before opening a node editor, write one card per shot:

The card prevents a prompt from mixing story, camera, physics, and finishing instructions. It also gives a reviewer a direct answer when a clip looks attractive but fails the shot’s actual purpose.

For continuity, track what the viewer should perceive across shots: screen direction, subject scale, light direction, wardrobe, product orientation, and background landmarks. The Academy of Motion Picture Arts and Sciences’ overview of continuity is a useful reminder that screen storytelling depends on deliberate visual relationships, not isolated frames.

Map dependencies around the shot, not the provider

Use named responsibilities rather than a maze of model nodes. The graph should answer “what does this shot depend on?” at a glance:

1. Shot brief: action, camera, duration, aspect ratio, and acceptance test.
2. Reference frame: the approved start image or frame.
3. Motion candidate: one action with one camera decision.
4. Continuity review: first, middle, and final-frame observations.
5. Trim or handoff: the accepted moment that feeds the next shot.

Add Character, Product, Style, or Background references only when they answer a specific continuity question. A second reference should explain why it exists: “preserve label geometry,” “match wardrobe,” or “continue warm side light.” If a reference does not change the review decision, leave it out and keep the shot card simpler.

For example, a product reveal may have a Brief — reveal, Start — bottle-v3, Motion — place and rotate, and Review — label and endpoint path. If the next shot is a close-up, connect the accepted final frame to a new shot card rather than copying the entire first graph. That keeps the handoff explicit and prevents an unrelated lighting experiment from changing the product source.

Infiknit supports a visual canvas where nodes, groups, titles, viewport state, and durable media references remain connected. Image and video nodes can become inputs to later visual work, and eligible video operations can run through dependency relationships. Keep the model, duration, resolution, and reference choices visible on the branch so the next shot starts from evidence rather than a guessed prompt.

Read Build node-based AI image generation graphs for still-image dependencies and Connect generation nodes on a visual AI canvas for connection decisions.

Motion and continuity are different tests

Motion asks whether the action is plausible: does the hand move toward the product, does the camera push in smoothly, does the object stop where requested? Continuity asks whether the shot belongs beside its neighbours: is the product facing the same way, does the light come from the same side, and does the final frame make the next cut possible?

Review them separately. A beautiful motion candidate can still be unusable if its last frame cannot hand off. Write the failure in concrete terms—“label turns 30 degrees away by frame 48”—rather than “identity drift.” The next branch can then change the source frame, motion strength, duration, or camera instruction intentionally.

A practical shot-graph method

1. Prove the hardest shot first. Pick the moment most likely to fail: a hand interaction, product rotation, face turn, or match cut. Use a short duration and a clean source.

2. Freeze identity before movement. Check product geometry, labels, face, wardrobe, and background landmarks in the start frame. Do not add a complex camera move while identity is still uncertain.

3. Choose one motion. “Hand lifts the box and stops at chest height” is testable. “Make it cinematic” is not. Keep camera movement separate from subject movement.

4. Inspect three frames. Check the first, middle, and last. Look for hands, text, object geometry, temporal warping, and the end condition.

5. Branch deliberately. Change one variable—action, lens, speed, reference, or model. Name the branch after that change.

6. Trim accepted material. Trim from a verified moment, preserve the original candidate, and record the in/out points. A trim should improve the handoff, not hide a failed setup.

7. Hand off with context. Link the accepted final frame or clip to the next shot, include the continuity note, and carry over the settings that matter.

Handoff patterns for common sequences

Match cut: end shot one with a stable product orientation; use that frame as the start reference for shot two. Compare silhouette and screen position before adding motion.

Action continuation: trim at the moment the hand reaches the product, then begin shot two from that frame. Preserve the direction and speed of the action.

Cutaway: use a detail shot only when it supports the same claim. Record which product or character feature the cutaway must preserve.

Dialogue or voiceover: keep audio timing separate from the visual graph. Pass the accepted visual duration to the audio brief rather than stretching a shot after narration is approved. W3C’s prerecorded-media caption guidance supports treating captions and timing as first-class review requirements.

Continuity checklist

At every handoff, check:

• subject identity and product geometry;
• screen direction and camera height;
• scale and eyeline;
• lighting direction, colour, and shadow shape;
• wardrobe, props, and background landmarks;
• frame rate, duration, crop, and safe area;
• the exact first frame and last frame;
• whether captions or audio need a new timing note.

Save the observation beside the branch. “Pass: label readable, hand exits right, next shot starts on same angle” is useful evidence. “Looks good” is not.

Plan for the edit, not only the generation

A generated clip is raw material for an edit. Mark where a cut can occur, whether the shot needs handles before and after the action, and whether the next shot begins on motion or on a held frame. These notes prevent a technically successful generation from becoming unusable in the timeline.

For a short vertical ad, preserve a clean opening frame for the first-second hook and a readable final frame for the call to action. For a tutorial, preserve enough setup that a viewer can tell what changed. In both cases, carry the crop and safe-area decision through the handoff; a 16:9 source may lose the very detail that makes a 9:16 cut work.

Infiknit boundaries that affect graph design

The canvas can keep working evidence connected and can queue supported image and video operations through validated tools. It does not turn every node into a queue target, remove provider-specific input limits, or guarantee that a generated clip will preserve identity. Cycles and unsupported connections should be treated as design feedback: return to the last accepted node and create a legal branch.

Do not represent an uncompleted provider action as an accepted handoff. A visible output, its source, and a review note are the proof that a shot is ready for the next dependency. This is especially important when a long sequence is built from several providers with different duration and reference rules.

Failure modes and fixes

The first frame matches but the last frame drifts. Shorten the shot, simplify motion, or branch from a stronger reference. Do not hand off the drifting frame.

The hand or prop changes shape. Reduce the action to one movement, use a cleaner source, and inspect the middle frame before adding camera movement.

The next shot jumps in scale. Record the final subject size and camera height, then match them in the next shot’s start-frame brief.

A trim removes necessary setup. Restore the preceding beat or use a new start frame that already contains the intended state.

The graph becomes unreadable. Group by scene, keep one branch per meaningful test, and archive failed experiments that no longer teach a boundary.

Captions no longer fit after a trim. Recalculate the transcript timing and review the new crop at phone size before export.

A shot review worksheet

For every candidate record source frame, shot purpose, camera intent, subject action, fixed details, duration, aspect ratio, model and settings, first-frame result, middle-frame result, final-frame result, and reviewer decision. This creates a compact record that can travel with an export. It also makes a failed branch useful: “label leaves frame during pan” points to a different repair than “endpoint has no caption space.”

Review a sequence in order after reviewing shots individually. Check screen direction, scale, light direction, background, character or product identity, and the transition between endpoints. If the sequence needs a bridge, create a new shot brief and reference it to the accepted endpoint. Do not force a continuation when a clean cut is clearer.

For handoff, keep the accepted frame and its source connected, include the intended next operation, and record which branches were rejected. A second creator should not have to inspect every render to find the approved take. The graph is successful when it reduces that search and keeps one revision from disturbing unrelated shots.

FAQ

How many nodes should a video shot use?

Start with a brief, reference, motion candidate, review, and handoff. Add nodes only for a real source, variable, or approval gate. Five legible nodes are better than a large graph that hides the decision chain.

Should I extend a failed clip?

No. Fix the first failing requirement and generate a new short candidate. Extending a clip with a broken identity or continuity usually carries the defect into the next shot.

What belongs in a video handoff?

Include the accepted frame or clip, in/out points, shot purpose, continuity observation, aspect ratio, duration, model and settings, and the next shot’s acceptance condition.

Can a node graph guarantee continuity?

No. It can make continuity requirements explicit and preserve evidence. The rendered frames still need human review because provider outputs vary.