Chapter 10 of 10. Prerequisites: Chapter 7 and Chapter 8. Chapter 3 taught you the constraint: first-frame chaining and reference media are mutually exclusive, so a shot cannot start on the previous shot's last frame while keeping your character assets and motion ref. This chapter is what that constraint looks like after another production went straight at it. The answer is not a workaround in the API. It is a technique, and every lever in it is load-bearing.
The problem, restated with money on it
Chapter 3's constraints catalogue includes the 400 you hit when you try to be clever: first and last frame content cannot be mixed with reference media content. In practice that means each render in a multi-shot sequence re-establishes the characters from the assets, in a neutral pose, wherever the motion ref happens to start. Cut two of those together and the audience sees a reset: the clinch that ended shot one has evaporated, the fighters are square again, sometimes they have even swapped sides. Chapter 8 taught you to mask resets with framing variety and impact flashes. That works for coverage. It does not work when the choreography genuinely continues across the cut, and a fight is mostly choreography that continues.
The production that cracked this was a full recreation of a six-minute stick-figure fight, rebuilt shot by shot with two registered character assets and a scene plate. The continuity requirement was not optional there. What came out of it is a repeatable chain.
The weld chain
Three levers, all required. Drop any one and the weld fails in a specific, observable way.
- Render a long continuous take
One render, up to the 15-second ceiling, covering a contiguous run of the motion ref. Within a single take, continuity is free: the model holds the characters, the scene plate holds the world, and the choreography flows because it never left one generation. The take's END is what matters next, so prefer source runs that end in a clear, separated pose. A take that ends in a tight tangle of bodies cannot be welded, for reasons you will see in lever two.
- Line-art the final frame as the next shot's anchor
Extract the take's last frame and run it through the same edgedetect line-art pass chapter 5 taught for motion refs. Prepend that line-art frame to the NEXT segment's motion reference. A photoreal frame gets blocked by the real-face detector, and a face-occluded photoreal frame passes moderation but fails to anchor. The abstract line-art contour is the form that both clears moderation and actually steers the model: the next render opens on the prior pose. This only works when that pose is clear and separated. A busy close-up reads as noise, and the model ignores it.
- Render a SHORT continuation whose prompt describes the transition
The continuation render is 5 to 6 seconds, not more. At 12 seconds the model drifts off the anchor and invents. And its prompt must describe the TRANSITION in words: what physically happens between the held pose and the new action. The production's first weld attempt reused a generic prompt and the join jumped. The retry prompted the break itself, release the clinch, separate, trade strikes, and the continuation opened still inside the clinch, held it, then broke exactly as described. The jump became a deliberate physical break. Same anchor, same footage, different prompt.
Then weld the two renders in the edit, and the continuation's end becomes the next anchor. Long take, line-art the end, short prompted continuation, repeat. A 27-second continuous fight shipped this way, from takes that the API insists cannot share a frame.
Why each lever is load-bearing
The chain was not designed. It was arrived at by watching each shortcut fail.
Skip the clear end-pose and the anchor is unreadable. The first weld test ended take one on a tight close-up clinch. The next render opened wide, in a neutral square-up. The model did not refuse the anchor; it just found nothing usable in a frame that was mostly overlapping bodies. Anchors weld on separated silhouettes.
Skip the line-art and you either get blocked or get nothing. Photoreal anchor frames hit the same real-face moderation wall as photoreal motion refs. Occluding the faces passes the filter but the anchor stops steering. The line-art contour is not a compromise between those two. It is the only representation tested that both passes and works.
Skip the transition prompt and the weld becomes a jump. This is the chapter 7 alignment rule operating at the seam: the prompt must describe what the footage should do, and at a weld the footage must do something the motion ref alone cannot specify, namely leave the previous take's pose believably. Prompt the transition and the model performs it.
Let the continuation run long and it forgets. Five to six seconds holds the anchor. Twelve drifts. The anchor is an initial condition, not a leash.
The discipline that guards the chain
The weld chain multiplied renders, and multiplied renders exposed a process failure: prompts were being reused across segments. The production elevated the fix to a constitution rule with enforcement. Every segment gets its own prompt, tailored to its action and its transition, with the character loadout pinned, and the pinning stays light: state what must stay true, exactly one katana each, never disappearing, and let the motion ref carry the choreography. Over-tailored prompts made the model fight its own reference.
The enforcement is worth copying even solo: a prompt ledger. The CLI appends every submitted prompt to a local ledger file and refuses to submit a render whose prompt is byte-identical to a previous segment's, before any money is spent. A rule you have never seen block anything is a preference. The ledger made it a gate.
What the finished chain looks like
The production this chapter documents shipped as a thirty-second graded fight: a welded 27-second spine, one single-frame limb-duplication artifact removed with an invisible one-frame drop, a grade, restrained handheld shake, a scored mix where the music sits at 0.92 against the fight's own action bed at 0.70 so the hits punch through, and branding. Total model spend for the whole discovery arc, including every failed weld experiment described above, landed around twenty-five dollars across two sessions. The technique transfers: any choreography-continuous sequence, fights, dance, sport, parkour, welds the same way.
Weld two takes of your own choreography
Take a motion reference with a genuinely continuous action run and produce a welded pair: one long take ending in a clear separated pose, then a 5 to 6 second continuation anchored on that pose's line-art frame, with a prompt that describes the physical transition in words. Use your chapter 3 CLI and your registered assets.
Expected behaviour
- The long take ends on a clear, separated pose you chose deliberately from the source run
- The continuation's motion ref has the line-art anchor frame prepended, and the render opens on the prior pose rather than a neutral reset
- The continuation prompt names the transition, and the join reads as one continuous action when the two renders are cut together
- A deliberately generic-prompt control render of the same continuation shows the jump the transition prompt prevents
- Your prompt ledger, or a manual log, shows a distinct prompt per segment
PROVE IT Cut the welded pair and the generic-prompt control into a side-by-side and watch the seam in both. If you cannot see the difference at the join, the weld was not the hard part of your source run; pick an end-pose mid-action and run it again.
Audit your renders for the audio default
Check every render in your current project for baked generated audio, strip it where present, and patch your CLI so audio generation is off by default for reference-media renders.
Expected behaviour
- A one-line probe (ffprobe) lists which of your existing renders carry an audio stream you never asked for
- Stripped copies exist with video untouched, using stream copy rather than a re-encode
- Your CLI now submits with audio generation explicitly disabled unless a flag opts in
- One render that previously failed with a sensitive-audio error resubmits clean with audio off, or you can explain why none of yours ever hit it
PROVE IT Show the ffprobe output before and after for one render, and the CLI diff. A default you have never overridden is a decision the vendor made for you.
Your welded continuation opens in a neutral wide shot instead of the previous take's pose. Which lever most likely failed?
What is the safe duration range, in seconds, for a welded continuation before it drifts off its anchor?
The first weld attempt and its successful retry used the same anchor and the same footage. Explain what changed and why it worked.
Show answer
Only the prompt changed. The failed attempt reused a generic prompt, so the model had no instruction for leaving the held pose and the join read as a jump. The retry described the transition itself, release the clinch, separate, trade strikes, so the continuation opened inside the clinch and performed the break as directed. At a weld, the motion ref cannot specify how to leave the previous take's pose; only the prompt can, and it must describe that transition in words.
↺ re-read: “The weld chain”