Clean the handoff frame before you chain clips

Chaining clips is how you get past a model's duration ceiling: take the last frame of clip one, feed it in as the opening frame of clip two, repeat. The mechanic is sound and it is what everyone does. What most people skip is the step in the middle, and skipping it is why chains that look fine at link two look wrong by link five.

The last frame of a generated clip is, structurally, the dirtiest frame in it. Handing it forward unmodified is why chains degrade — and the fix is a single intermediate pass on one still.

Why error concentrates at the end of a clip

A video generation is not a set of independent images. Each frame is produced with reference to what came before, so small errors do not stay where they started — they are inherited by the next frame and become part of the context the frame after that is built on. Whatever the model got slightly wrong at frame ten is baseline truth by frame sixty.

So the final frame is where accumulated error is highest, and it is exactly the frame the chaining workflow selects. Feed it forward and the next clip starts from the previous clip's accumulated error, then adds its own. Chain degradation compounds rather than staying put, and by the fourth or fifth link you are generating from an input that has drifted well away from anything you approved.

Capping clips at three to five seconds is the standard mitigation for drift inside a clip. Cleaning the handoff frame is the mitigation for drift between them, and you need both.

The intermediate step: re-mint the frame

The step is: extract the last frame, run it through image-to-image as a still, and use the cleaned output — not the raw extract — as the seed for the next clip.

What this does is give a still-image model a chance to resolve what the video model left ambiguous. Video models spend their capacity on motion and temporal coherence; a dedicated image pass spends all of it on one frame. It tightens edges, resolves a hand that was mid-blur, and clears the small structural noise video generation leaves behind.

The settings that matter:

Keep denoising strength low. You are cleaning a frame, not reimagining it. High denoise gives you a beautiful image of a slightly different scene, which produces a visible jump at the join.
Prompt the frame, not the shot. Describe what is in the still: subject, wardrobe, set, lighting. Do not carry over motion language from the clip prompt; there is no motion in a still and the model will try to imply some.
Change nothing compositionally. No reframing, no crop, no aspect change, or the next clip inherits a jump cut.
Re-anchor to your reference. If the sequence has a locked character reference, supply it here too. This is the cheapest chance to pull identity back toward the anchor before it gets baked into another clip.

Upscaling is a reasonable alternative when the frame is structurally fine but soft, and a poor substitute when it has real errors — upscaling makes existing detail bigger, including detail you wanted removed. Clean first, enlarge second. The same principle governs finishing a whole sequence: deflicker before you step up resolution, never after.

How to tell a frame is clean

Judge the still at 100 percent, not fitted to a window. Reduced-size review passes almost everything.

Work through the frame in this order — roughly most-likely-broken to least:

Hands and fingers. Count them. Check that each one terminates in a fingertip rather than merging into an adjacent finger or into whatever is being held.
The subject's silhouette. Trace the outline against the background. Halos, smeared edges and a boundary that goes soft in one place and hard in another are all inherited artifacts.
Hair and fabric boundaries. These are where temporal coherence degrades first, so they are where a last frame carries the most residue. Look for strands that dissolve into the background and folds that do not resolve into a consistent surface.
Small repeated geometry. Railings, buttons, keyboard keys, window mullions, spokes. Repetition is where a model averages, and averaged repetition produces spacing that is subtly wrong.
Text and signage. Any legible glyph in the frame either reads correctly or it does not. Models learn glyph shapes rather than letterforms, so background text is usually the first thing to go.
Gradients. Skies, walls, out-of-focus backgrounds. Banding and blotching here are quiet and they enlarge badly.
The background props. The thing on the shelf behind the subject. Nobody checks it, and it is often where the frame has quietly changed since the clip's opening.

A frame passes when nothing on that list would embarrass you as a standalone photograph. That is a higher bar than "looks fine in playback," and it is the right bar, because the next clip builds dozens of frames out of it and every flaw gets amplified rather than averaged away.

If a frame fails on two or three of those points, do not rescue it. Pull a different frame — three or four frames earlier is usually still on the same beat and often cleaner — and trim the tail of the clip to match. At 25 fps that is well under two tenths of a second, which nobody will miss, and it saves the chain.

When re-minting is the wrong tool

Three situations where the step does more harm than good.

When the frame is mid-motion. A frame captured mid-gesture or mid-stride contains genuine motion blur, and an image pass interprets that blur as something to resolve. You get a sharp frame that no longer sits inside the motion arc, and the next clip starts from a pose that does not connect. Chain on frames where the subject is momentarily settled — end of a gesture, top of a step, a held look.

When identity is already fragile. An image pass at anything beyond very low denoise can move a face. If the sequence is already drifting, re-minting without a reference anchor is as likely to push it further as to pull it back. Anchor or skip.

When the shot's end state is already decided. If you know exactly where the next clip should land, first-last-frame generation beats a cleaned handoff: you supply both endpoints and the model solves only the motion in between, a bounded problem rather than an open one. Building a whole scene that way is covered in keyframe chaining for longer scenes, and the trade-off against plain extension in extend, chain, or first-last-frame.

The order of operations for a chain

Generate clip n, capped at a length where the model still holds — three to five seconds is the working range for most current video models.
Review clip n at full resolution before you extract anything. A clip that fails now will fail worse three links later.
Extract the handoff frame. The editor can turn video into still photos so you are working with a real frame rather than a screenshot of a player, and pulling stills from footage you already rendered covers the wider uses of that.
Run the frame checklist above. Reject and re-pick if it fails.
Re-mint the frame in image-to-image at low denoise, with the character reference attached. The AI photo editor covers the still-image side of this.
Seed clip n+1 with the cleaned frame, not the raw one.
At assembly, colour-match every clip to its neighbours in the timeline. Exposure and colour drift across links is expected and is not a prompting failure — it is a grading job, and it is not optional on a multi-clip sequence.
Deflicker and clean before any upscale pass, then step up resolution.

Step seven gets left out most often and it takes almost no time. Even a well-chained sequence arrives with slightly different exposure per link, because each generation made its own decisions about the light. Matching them in the video editor at assembly is what makes the chain read as one continuous scene. The editor's 480p preview pass is free and carries a short per-user cooldown, so you can check the assembly's rhythm before committing; the final export is charged once regardless of clip count.

FAQ

How many links can a chain survive?

There is no fixed number, and anyone quoting one is describing their footage rather than yours. What is reliable is the shape of the failure: unmanaged chains degrade faster with each link, and cleaning the handoff frame plus re-anchoring to a reference flattens that curve considerably. Review every link rather than budgeting for a count.

Does this apply to native extension features too?

Less so. When a model extends a clip internally it usually has access to more of its own state than a single exported frame carries, so there is no dirty handoff to clean. The re-mint step is for chains you assemble yourself, where one image is the only thing crossing the boundary.

What denoise level should I use on the re-mint?

Low enough that the output is recognisably the same frame when you flip between them. There is no universal number — it depends on the model and how noisy the input is — so calibrate by comparison. If you can see the composition change, you have gone too far.

Can I just upscale the last frame instead?

Only when the problem is softness rather than structure. Upscaling faithfully enlarges what is there, so a frame with a fused finger comes back with a larger fused finger. Run the checklist first: structural failures need a re-mint or a different frame.