The Real Cost of a Reshoot in a Generated Pipeline
A client's small tweak and a client's swap-the-look request aren't the same change order. A credit taxonomy for what each revision type actually costs.
Anyone who's produced live-action learns the instinct fast: a continuity fix costs a fraction of a shoot day, and swapping the location costs almost as much as reshooting, because half the crew logistics repeat regardless of how small the note sounds. That instinct is genuinely useful and it does not transfer cleanly to a generated pipeline. The four things a client might call "just a small change" price four completely different ways in credits, and the tier that looks smallest on the brief isn't always the one that's cheapest to execute.
A taxonomy of four revision types
Retake is the narrowest scope available and the cheapest: regenerate one segment of an existing clip, bounded by a start time and a duration, everything outside that window untouched. LTX 2.3 Retake Video's own description gives the clearest sense of what this actually covers — a source video plus a prompt describing the desired change, its own example being something as small as "change flower to red rose" — at 10 credits per call. This is the continuity fix: one detail, one bounded window, nothing else in the frame or timeline at risk.
Video-to-video edit widens the scope from a bounded segment to the whole clip, applying a described, prompt-based modification across it rather than confining the change to a window. Kling O1 Standard and Kling O3 Standard both price this at 13 credits; the non-Standard Kling O1 tier runs 17. Still a targeted change, but no longer fenced to a few seconds — closer to "re-grade this whole shot" than "fix this one frame."
Video-to-video reference is where the jump actually lands, and it's a bigger one than the name suggests. Kling O3 Standard's own description names it directly: "reference-based editing for style transfer," at 63 credits — and the Pro tier, described as reference-based editing "for style and character transfer," runs 84. This tier isn't editing the clip in isolation; it's holding a second input — the reference — in play across the entire generation, which is a meaningfully heavier job than modifying what's already there.
Full regeneration is the reshoot: no preservation guarantee at all, priced as a first-class generation rather than a modification of one. Kling Video V3 4K starts at 210 credits for the base clip and scales to as much as 630 for a longer render. Nothing from the original survives except whatever the new prompt happens to reproduce.
Why the jump between tiers isn't linear
Retake and edit sit close together — roughly a one-and-a-half-times gap between the cheapest and most expensive options in that pair — because both are constraining the model to touch less of the frame or the timeline than a fresh generation would. Reference jumps far past both, four to six times the edit tier's price, because a reference call isn't a lighter version of an edit — it's carrying an entire second input through generation alongside the first, which costs closer to what a fresh generation costs than what a targeted patch does. Full regeneration is another two-to-three-times step past that, because at that tier nothing is being held constant at all.
That's not a smooth curve. It's three tight clusters with two real cliffs between them, and the cliffs are exactly where a client's or a producer's live-action instinct gets it wrong.
Put a week of revision requests through the taxonomy and the gap compounds fast. Three small continuity notes as retakes run about 30 credits total. One broader "re-grade this whole shot" note as an edit adds roughly 13 to 17 more. A single "match our new campaign's look" note misread as an edit instead of a reference undercharges by 50 credits or more against what the job actually does — and if that note comes up twice in a revision round, the misread alone is worth more than every retake on the list combined. The taxonomy isn't pedantry for its own sake; it's the difference between a revision round that stays predictable and one where the tenth request quietly costs as much as the first nine put together.
Where the instinct actually misleads
The failure mode isn't misjudging retake or full regeneration — those two are intuitive at both ends, cheap-and-narrow versus expensive-and-total. It's the middle. "Video to video" reads as one category to someone pricing off vibes, and it isn't: edit and reference are both "not a full regeneration," both keep the source clip in the request, and they price roughly five times apart. A request that sounds like "keep everything, just change the outfit's style to match our new campaign" is a reference call, not an edit call, the moment it's asking the model to transfer a look from somewhere else rather than modify what's already in the frame — and treating it like a 13-credit edit because it "isn't a full reshoot" is exactly the assumption that gets the estimate wrong before the job even starts.
The genuinely useful part of the live-action instinct still holds, once it's pointed at the right axis. A live-action swap that touches blocking or geography costs nearly as much as a reshoot because it drags real logistics back into scope; a reference-based generation costs nearly as much as a fresh clip because it drags a second full input back into scope. Same underlying shape, different mechanism.
Pricing the order before you commit to it
None of this has to be estimated from memory. estimate_cost takes a planned set of generations — model, count, duration — and returns the credit total against the current balance before anything dispatches, the same pricing logic the real charge uses. The discipline worth building around a client change request is simple: never let "it's just a small tweak" go straight to execution. Price the actual tier first, because the difference between what a request sounds like and what it actually calls for is precisely the ten-to-eighty-plus credit gap this taxonomy maps out. Versely's quality-per-credit report and the full pricing page are worth having open next to that conversation, since the right tier for a given note isn't always obvious from the words a client used to describe it.
A Versely walkthrough
A client note like "keep the shot, just make it feel like our new brand look" is genuinely ambiguous between two tiers that price five times apart, which makes it exactly the kind of request worth pricing both ways before picking one:
"Estimate the cost of two options for this clip: a video-to-video edit applying our new color and mood direction as a prompt, versus a video-to-video reference using our new campaign's hero shot as the style source."
estimate_cost returns both totals — the edit tier at roughly 13-17 credits, the reference tier at 63-84 — against the current balance, before either one runs. If the edit tier's prompt-based description covers what the client actually meant, that's the honest cheaper read; if the request genuinely needs another asset's look pulled in, the reference tier is the correct spend rather than an edit call quietly under-delivering on it. Either way, checking credits before generating and Versely's redo-a-previous-generation flow turn "just a small change" into a priced decision instead of a guess that gets corrected after the fact.
That five-minute check is the whole discipline. It doesn't require memorizing every model's credit rate — it requires treating "how big is this change, really" as a question with a priced answer, asked before the job runs rather than reconstructed afterward when the revision round costs more than the plan.