Restate your subject descriptors **identically** between rounds — same words, same order, not improved. Hexcoded's own documentation names altered descriptors as a cause of drift through latent re-encoding. Then describe only what happens next. The source clip already supplies the scene, so re-describing it competes with footage the model can already see.
Extend is the one prompt field where writing something better makes the output worse, and the reason is mechanical rather than stylistic.
What Extend is conditioned on
Actual footage, which is what makes it different from everything else.
Create has only your words. Edit has a source clip plus your description of a change. Extend has a clip you already rendered, and appends to it — so space, light, wardrobe and identity all carry forward from real frames rather than being re-guessed from a description.
That's a much stronger starting position than a reference image. A reference says what a character looks like; a rendered clip says what this specific shot looks like, in motion, at this moment.
Which means your prompt has less work to do than in Create, and a different job.
Why identical wording matters
Because small variations compound through the chain.
Hexcoded's own model documentation names the mechanism: latent re-encoding. Each Extend round re-encodes the previous output, and altering subject descriptors between rounds introduces variation that accumulates rather than resolving.
So "woman in a charcoal suit" in round one, "woman in a grey suit" in round two, and "woman in a dark suit" in round three isn't three descriptions of the same person. It's three slightly different instructions layered onto a chain that was already carrying forward the first one.
The counter-intuitive part: improving the phrasing is worse than leaving it clumsy. If round one said something awkward, round two should say the same awkward thing. Consistency beats quality here, which is the opposite of every other prompting instruction you'll read.
If round one was clumsy, round two should be clumsy identically. Consistency beats quality in a chain.
What the prompt should actually contain
Three parts, in this order.
The subject descriptors, copied exactly
Paste them. Don't retype, don't tidy, don't improve. This is the part that must not vary, and pasting is the only reliable way to guarantee it.
What happens next
The continuation, and only that. One action, as in Create. The model already knows where everything is — you're telling it what changes from here.
Nothing about the existing scene
Framing, light, wardrobe, location. All of it is in the source clip. Describing it again gives the model a second opinion on something it can already see.
What that looks like
Round two of a three-round chain.
@maya, charcoal suit, short black hair. She lowers her hand from the window frame and takes one step back into the room.
Descriptors first, copied from round one. Then one action. Nothing about the kitchen, the light or the camera — all of that is in the source clip.
Note that the descriptors appear even though a handle is present. That's deliberate and it's the one place where the usual rule bends.
Elsewhere, a handle plus an inline description is a conflict. In a chain, the documentation's guidance is about descriptor consistency specifically — so if you included descriptors in round one, include the same ones in round two. If you included none, include none.
The rule is consistency within the chain, not presence or absence. Starting with descriptors and dropping them mid-chain is the same failure as changing them.
modelsWhich models let you extend at all
Seven of nineteen, and it's worth knowing before you plan a chain.
Seedance 2.5, Seedance 2.0, Seedance 2 Fast and Seedance 2 Mini — the whole family extends, which makes it the most chain-friendly group on the roster.
Veo 3.1 and Veo 3.1 Fast extend too. Roughly seven seconds per run, up to around twenty runs, with output matching the source clip. That's how an eight-second model reaches past a minute.
Grok Imagine 1.5 completes the list.
Everything else — Kling, Wan, MiniMax, HappyHorse, Hailuo, LTX, Vidu, Gemini Omni Flash and Veo 3.1 Lite — has no Extend mode. Generate a shot on one of those and the chain stops there.
What Extend inherits, and can't change
Two settings come from the source and the selector is hidden for both.
Aspect ratio. Decided by the first Create generation in the chain. Extend takes it and offers no choice.
Resolution. Same. So a chain started at 720p stays at 720p however many rounds you add.
Which makes the first generation in a chain the most consequential one. Framing and quality are both fixed there, and nothing downstream can revisit either.
- Restate subject descriptors identically between rounds. Paste them rather than retyping
- Improving the phrasing is worse than leaving it clumsy. Consistency beats quality in a chain
- Hexcoded's documentation names the mechanism: altered descriptors cause drift through latent re-encoding
- Describe only what happens next. The source clip already supplies framing, light, wardrobe and location
- The rule is consistency within the chain — dropping descriptors mid-chain fails the same way as changing them
- Seven of nineteen models extend. The Seedance family, Veo 3.1 and Fast, and Grok Imagine 1.5
- Veo extends roughly seven seconds per run, up to around twenty runs
- Aspect ratio and resolution are both inherited. The first generation in a chain fixes them permanently
Three things in order — your subject descriptors copied exactly from the previous round, then what happens next, and nothing about the existing scene. Framing, light and location all carry forward from the source clip, so describing them again competes with footage the model can already see.
Because each round re-encodes the previous output, and Hexcoded's documentation names altered subject descriptors as a cause of drift through latent re-encoding. Small variations don't resolve across rounds — they accumulate.
Keep it. Improving the phrasing mid-chain introduces exactly the variation you're trying to avoid. Consistency beats quality here, which is the opposite of normal prompting advice — fix the wording on the next chain rather than this one.
Match whatever round one did. If you included descriptors alongside the handle, include the same ones every round. If you included none, include none. The rule is consistency within the chain rather than presence or absence.
Seven of nineteen — Seedance 2.5, 2.0, 2 Fast and 2 Mini, plus Veo 3.1, Veo 3.1 Fast and Grok Imagine 1.5. Kling, Wan, MiniMax, HappyHorse, Hailuo, LTX, Vidu, Gemini Omni Flash and Veo 3.1 Lite have no Extend mode at all.
No. Extend inherits both aspect ratio and resolution from the source and hides the selectors. Whatever the first Create generation in the chain used is what every round gets, which makes that first generation the most consequential decision in the sequence.
Chain it, don't cut it
Extend appends to a clip you've already rendered, so space, light and identity carry forward from real frames. Seven models support it, and every setting is visible before you generate.
Open Creative StudioMore on prompt structure, per-tool syntax and common mistakes in Prompts.