Start Frame, End Frame: Directing the Bridge

Two stills, one bridge

The start/end keyframe technique is simple to state and easy to underuse: generate a high-resolution start still and a high-resolution end still — both produced deliberately through an image model, not pulled from a video frame — and feed both into the video model as the two anchor frames. The motion prompt then describes ONE continuous shot as a timestamped sequence of beats, closing with an explicit instruction that the clip must end in a locked still frame matching the end reference exactly. The model's job is to invent everything that happens between those two points.

It's tempting to think of this as simple interpolation — "the model just tweens between image A and image B." That undersells what's actually happening, and underusing it that way leaves the technique's real value on the table.

The model's actual superpower

Interpolation implies a smooth, predictable path between two nearly-similar states. The start/end technique is far more useful than that, because the two stills don't have to be similar at all — they can represent a genuine transformation. A character in shadow to a character in full light. A calm expression to a triumphant one. An empty street to the same street mid-transformation into something else entirely. The gap between the start and end stills is where reveal, arrival, and transformation beats live, and directing that gap deliberately — rather than treating it as an afterthought — is where this technique earns its place as a director's primary tool, not a convenience feature.

The discipline is to deliberately engineer that in-between moment into the shot, rather than picking two arbitrary stills and hoping something interesting happens between them. Write the motion prompt as a genuine timeline: what's true at the start, what shifts and when, what the climax beat looks like, and how it settles into the end frame. The model isn't filling a gap you left empty — it's executing a transformation you directed, frame by frame, even though you never generated the frames yourself.

Continuity across a cut: feed the prior final frame

Production doesn't happen one shot at a time in isolation — scenes chain together, and each cut is a place continuity can quietly break. The technique for holding a chain together: feed the PRIOR scene's actual final frame into the next scene's generation as an additional reference, alongside that scene's own start and end stills.

This gives the model something concrete to hold onto for composition and color across the cut, instead of reinventing the visual language of the world at every scene boundary. Without it, even a well-directed individual scene can feel disconnected from the one before it — subtly different color grading, a framing choice that doesn't rhyme with what came before. Feeding the actual prior frame closes that gap the same way the start/end technique closes the gap within a single shot.

Case study: from three stitched clips to one continuous shot

A scene built early in production illustrated the failure mode clearly. The plan had been to generate three separate five-second clips and stitch them together in the edit — the thinking being that three shorter, simpler generations would each be easier to get right than one longer, harder one.

On review, the result read as exactly what it was: three ugly cuts. Each clip had been generated independently, with its own interpretation of pacing, framing, and motion, and no amount of editing polish closed the seams between them. The fix wasn't a better transition — it was rewriting the scene entirely as one continuous 15-second shot, using a start still and an end still as anchors, and a single timestamped motion prompt spanning the full duration: a push-in beat, a reveal beat, a climax beat, and a dissolve beat, all written as one continuous instruction rather than three separate prompts. The result read as one shot because it was directed as one shot.

The lesson generalizes past this one case: whenever a sequence is starting to feel like "clip one, then clip two, then clip three," check whether it should instead be "one continuous shot, four beats." The start/end technique is very often the better tool, even when the instinct is to break a complex shot into simpler pieces.

Realistic cost

For planning purposes: a full clip generated this way typically runs somewhere in the range of 60-70 credits for a 15-second shot at the mid resolution tier (the same range you'll budget against in Credit Economics later in this track), while each still frame (start or end) costs well under 1 credit to generate on its own. That asymmetry is worth internalizing — the stills are cheap relative to the clip, which is exactly why the style-anchor discipline from the previous lesson pays for itself here specifically: get the start and end stills right, cheaply, before spending the larger amount on the clip itself.

Three formats, one underlying technique

The start/end bridge isn't a single-purpose trick — it's the primitive underneath three distinct delivery formats you'll use across this track and in real production.

Journey / continuous-shot is the technique used directly, for a single unbroken shot — the format best suited to reveal, arrival, and origin beats, exactly like the case study above. Spectacle montage chains several bridged shots together, cut around a power or transformation moment, and is covered in full in the next lesson on storyboard grids. Reference+prompt card is a delivery wrapper — a locked, teachable format — around either of the other two, and is covered fully in the capstone. Recognizing these as three compositions of one technique, rather than three separate skills to learn from scratch, is what lets you move fluidly between them once the underlying bridge technique is solid.

Build it: assemble the bridge prompt

You're going to implement the function that turns a scene brief — a start description, an end description, a duration, and a beat list — into the correctly formatted motion prompt, including the mandatory closing instruction that locks the clip onto the end frame.