Multi-Scene Continuity in AI Video: Prompts with Reference Frames and Positive Locks

AI video creator Yurii Yeltsov (@yura_elkin) has unveiled a structured prompt architecture designed to maintain character and environmental continuity across a dramatic narrative twist and a hard scene cut within generative video models. Demonstrating the technique through a satirical twist on the Adam and Eve narrative—dubbed "The Chinese Edition," in which the pair roasts and feasts on the serpent rather than succumbing to temptation—Yeltsov details a reliable framework for directing multi-scene progression, camera framing, and state changes within a single generation pass.

Image source: @yura_elkin / X

Maintaining character identity, clothing consistency, and object permanence across scene transitions remains one of the most frustrating bottlenecks in generative AI filmmaking. While models frequently deliver convincing motion within a continuous single shot, introducing a camera cut or altering an object's state often triggers visual drift, hallucinated wardrobes, or logical contradictions. Yeltsov's method counters this instability by combining image reference frame anchoring, explicit hard cuts, second-by-second timeline direction, and an explicit 'POSITIVE LOCKS' block that separates immutable features from narrative state changes.

The Continuity Challenge in Generative Multi-Scene AI Video

Directing multi-scene short-form content in generative video models introduces three core technical hurdles:

• Cross-Shot Character and Wardrobe Drift: The facial geometry and clothing established in the opening shot frequently warp or reset into unrelated styles once a camera cut occurs.
• Narrative Object State Mutation: Directing a living subject to transform into another logical state—such as a live serpent becoming roasted food on a wooden spit—often confuses the diffusion process, yielding surreal hybrid artifacts rather than a coherent cause-and-effect transition.
• Floating Subtitle and Lip-Sync Detachment: Generating on-screen dialogue frequently results in illegible, melting glyphs along the bottom edge of the frame rather than synchronized comic-style speech bubbles.

To overcome these failure modes, Yeltsov approaches prompt engineering not as descriptive prose, but as a modular cinematic shooting script. By anchoring each shot with dedicated reference frame keys and decomposing the 15-second duration into granular second-by-second actions, the prompt gives the generation model unambiguous temporal instructions.

Five Structural Pillars of the Multi-Scene Prompt Architecture

The 15-second multi-scene production prompt is structured into five distinct operational blocks:

1. Dual Opening Frame Reference Anchoring

Rather than relying on generic text descriptions to initialize each shot, the prompt explicitly binds reference images to scene starting points: 2uekxqm3gz serves as the opening frame for Scene 1, and dp7h65n1n4 anchors the opening frame for Scene 2. This prevents the model from hallucinating a new setting or character appearance immediately after the cut.

2. Global Scene Context and Explicit Cut Directing

The prompt establishes an overarching photorealistic cinematic style, sets the duration to an exact 15 seconds, and explicitly commands a hard cut between them. Defining a hard cut forces the model to execute an immediate visual transition at the specified timestamp rather than attempting a gradual, distorted morph.

3. Floating Comic-Style Text Rules

The typography instructions disallow conventional bottom subtitles, mandating large, bold-outlined comic book lettering that pops directly adjacent to the active character. Imposing a minimum display duration of 1.5 seconds guarantees legibility and anchors viewer focus during the dialogue delivery.

4. Timestamped Storyboard Direction

The narrative timeline is divided into Scene 1 (00:00–00:07, "The Offer") and Scene 2 (00:07–00:15, "The Feast"). Specific camera framing—such as close-ups on the serpent's face, wide shots establishing the garden, and tight shots on the chopsticks and roasted meat—is pinned to exact timestamps, steering the model's pacing and facial expressions.

5. Positive Locks and Invariant State Control

The POSITIVE LOCKS block functions as an invariant enforcement layer. It explicitly commands the model to preserve immutable features (identical leaf outfits and characters across both scenes) while strictly isolating mutable narrative objects: Scene 1 requires a living serpent with iridescent scales holding an apple, whereas Scene 2 strictly commands a dead, roasted serpent on a spit with an explicit NOT alive constraint.

Full Verbatim Production Prompt and Director's Notes

The complete, verbatim prompt shared by Yurii Yeltsov is reproduced below:

2uekxqm3gz as opening frame Scene 1. dp7h65n1n4 as opening frame Scene 2.

SCENE CONTEXT: Photorealistic cinematic — two scenes with hard cut between them — 15 seconds total. Garden of Eden setting. Characters speak out loud with subtitles as floating comic-style text.

ALL ON-SCREEN TEXT: comic book style — bold outlined letters — large — clearly readable — floating near character — NOT bottom subtitles — visible 1.5 seconds minimum.

SCENE 1, 00:00–00:07 — THE OFFER:
00:00 — EXACT ref1 opening frame — Adam and Eve in leaf outfits standing in garden — serpent coiled on branch above — golden light — lush paradise.
00:01 — Serpent slowly extends toward them holding red apple — head lowering to their level — intelligent eyes — CLOSE UP serpent face — speaks out loud in tempting voice — comic text POPS bold near serpent head: "EAT THIS" — stays visible 1.5 seconds.
00:03 — Adam and Eve look at each other — pause — then both turn back to serpent — WIDE showing all three — garden background — golden light.
00:04 — Both SMILE widely — warm happy smiles — nodding slowly — Eve reaches toward apple — Adam leans in curious — CLOSE UP their smiling faces — satisfied nodding.
00:06 — CLOSE UP red apple in serpent tail — Adam and Eve hands reaching toward it — warm golden light — garden paradise around them. HARD CUT.

SCENE 2, 00:07–00:15 — THE FEAST:
00:07 — EXACT ref2 opening frame — Adam and Eve sitting at campfire — DEAD roasted serpent on wooden spit — golden brown — crackling — evening garden — same leaf outfits.
00:08 — Eve turns spit with both hands — practical satisfied expression — steam rising from roasted serpent — campfire orange light on faces.
00:09 — Adam picks up chopsticks — leans toward roasted serpent — CLOSE UP chopsticks pulling piece of roasted meat — steam — golden crispy skin.
00:10 — WIDE showing full scene — campfire — roasted serpent — Adam and Eve — dark garden evening background — peaceful and satisfied.
00:11 — Adam takes bite — chews — eyes closing in satisfaction — CLOSE UP his face — pure enjoyment.
00:12 — Adam looks at camera — grins — comic text POPS bold near Adam: "BETTER THAN APPLE" — stays visible — Eve laughs beside him — both laughing.
00:14 — WIDE final shot — both laughing at campfire — roasted serpent between them — garden night sky above — stars visible. FREEZE. End.

LIGHTING: Scene 1 — warm golden hour sunlight — paradise bright. Scene 2 — warm orange campfire only — dark garden background — intimate evening.

POSITIVE LOCKS: ref1 exact opening Scene 1 — leaf outfits on both characters. ref2 exact opening Scene 2 — same leaf outfits — same characters. Serpent in Scene 1 — alive — iridescent scales — holding apple. Serpent in Scene 2 — DEAD — roasted — golden brown — on spit — NOT alive. "EAT THIS" comic text near serpent Scene 1. "BETTER THAN APPLE" comic text near Adam Scene 2. Both characters speak out loud matching text. HARD CUT between scenes at 00:07. No watermarks.

Original source

• Yurii Yeltsov (@yura_elkin) Official Post on X: https://x.com/yura_elkin/status/2107761835512381524
• Detailed Prompt Thread on X (@yura_elkin): https://x.com/yura_elkin/status/2107762145693733269