One Character, Five Scenes: Creating a Consistent Story With AI Video

Yes, you can make a five-scene AI video with one consistent character from start to finish, and the method is simpler than the results suggest: build one character reference sheet, then generate every scene with a reference-to-video model using those exact same images. Kling 4.0, Google Veo 3.1, and Runway Gen-4 all animate a scene around your character’s photo instead of inventing a new face per run. Chain the clips with first- and last-frame transitions, add lip-synced dialogue where the story needs it, and you have a short film rather than five strangers in similar lighting.

The workflow still breaks in predictable places, and predictable means fixable.

How reference-to-video keeps one character consistent

Reference-to-video is a generation mode where you upload still images of your character (sometimes props and settings too) and the model animates a scene around them, rather than generating a person from text alone. The reference images act as anchors: the model matches facial structure, hair, and clothing to the images instead of rolling new dice for every clip.

Every major tool has a version under a different name: Kling’s Omni Reference, Google’s Ingredients, Runway’s References, Vidu’s Reference to Video, Hailuo’s Subject Reference. The mechanics rhyme: one to ten stills go in, one recognizable character comes out.

Two things changed between 2025 and late 2026. Reference capacity grew (Kling 4.0’s release page lists up to 10 reference images plus 5 video clips in one generation) and audio got folded in, so dialogue and lip movement generate in the same pass. Consistency is now a listed feature, not a prompt-engineering hobby.

Which AI video tool should you use for a one-character story?

Kling 4.0 is the most complete pick for a five-scene character story: the deepest reference support, up to 10 keyframes, 30-second single-pass clips, and native lip-synced dialogue. Veo 3.1 is the easiest if your story runs on dialogue, because its 8-second clips come with generated voices and sound. Runway fits image-first creators who want editing tools around the model.

Here’s the current lineup, with numbers from the vendors’ own pages (checked November 2026):

Tool (vendor)	Character reference features	Single-clip length	Entry cost
Kling 4.0 (Kuaishou)	Omni Reference: up to 10 images + 5 videos (15 assets), 10 keyframes, 3 voice profiles; native lip sync	3-30 s	From $10/mo (660 credits)
Veo 3.1 (Google)	Ingredients: up to 3 reference images; first/last-frame transitions; native dialogue audio	8 s	$0.40/s via API; Flow via Google AI Pro ($19.99/mo)
Gen-4 / Gen-4.5 (Runway)	References: one image holds a character, object, or location consistent	2-10 s (720p)	12 credits/s; Standard $12/mo (625 credits)
Ray3.2 (Luma AI)	Character reference; up to 16 keyframes per clip; start/end keyframe linking	Up to 20 s (1080p)	Plus $30/mo (10,000 credits)
Vidu	Reference to Video: 1-7 reference images	Model-dependent (Q2/Q3/Q4)	Free credits; paid from ~$8-10/mo

Fine print: Kling’s prices are list prices before regional promotions, and credit rates vary by model: the official Video 3.0 Omni guide lists 12 credits per second at 1080p with audio on. Veo’s $0.40 is the standard API tier with audio; Fast runs about $0.15 per second per OpenRouter’s listing. Runway’s Gen-4.5 outputs at 720p per its help center.

Also worth knowing: Hailuo’s Subject Reference has held characters steady from a single photo since January 2025, and Pika’s Character Studio locks a design you can summon into any clip. Ray3.2 is the keyframe specialist: the pick when your story is more choreography than dialogue.

Kling 4.0 accepts up to 15 combined reference assets (10 images plus 5 videos) and 10 keyframes inside a single 30-second generation. Kling 3.0’s ceiling was 7 images and 15 seconds.

Step 1: Define the character before you generate anything

Consistency starts as a writing problem, not a model problem. Before opening any tool, write a fixed identity block: age, build, hair color and style, eye color, skin tone, outfit from head to toe, and one memorable accessory: say, a brass compass on a cord. You’ll paste this exact wording into every single prompt, because wording drift causes appearance drift.

Then write the five beats: normal world, what the character wants, complication, crisis, resolution. A running example is a lighthouse keeper: dawn at the rail (1), a letter that changes everything (2), the storm rolls in (3), the light fails (4), morning after (5).

Keep the outfit identical across all five scenes if the story allows: costume changes are the most common drift trigger, and dodging them costs nothing.

Step 2: Build a character reference sheet

A character reference sheet is one image showing your character from several angles (face front-on, a profile, a three-quarter view, and a full body) on a plain background. It gives the video model something to match from every camera angle, not just the flattering one.

Make it with a strong image model: Google’s own Veo 3.1 announcement suggests Nano Banana Pro (Gemini 3 Pro Image) for ingredient images, and the community technique matches: generate the multi-angle sheet, check the face reads from the side, freeze it. Midjourney users can do the same with Omni Reference, which replaced --cref in V7.

Two practical rules. Keep the sheet’s background empty so the video model doesn’t import the room into every scene. Save the sheet and the identity block together as a kit: you’ll reuse both in every scene from here on.

Step 3: Generate the five scenes with the same references

This is the whole workflow, scene by scene:

1. Scene 1. Establish. Feed in the reference sheet plus a wide-shot prompt: the keeper at the rail at dawn, camera slow, no dialogue. Hold the shot for the full clip. This scene sets the visual contract: everything the audience will recognize later gets introduced here.
2. Scene 2. The want. Medium shot, interior, the keeper reading the letter. Same references; change only the framing and action. Keep lens and lighting language identical (“soft window light, 35mm”) across scenes so the model holds one visual tone.
3. Scene 3. The complication. The storm rolls in, and here you add dialogue if your story has any. Kling 4.0 generates lip-synced speech in one pass; Veo 3.1 generates voices natively. Write dialogue that fits inside one clip: roughly 15-20 words per 8-second Veo scene, and keep it to one speaker per clip.
4. Scene 4. The turn. The light fails. This is the motion-heavy scene, so generate two or three takes and keep the one where the face holds best. Motion is where drift shows first.
5. Scene 5. Resolve. Morning after. Mirror scene 1’s framing and palette words on purpose: the callback reads as closure, and reusing scene 1’s exact prompt structure exploits everything the model already got right.

Whatever the scene, the same three inputs go in: the reference sheet, the identity block, and your lens/lighting wording. Change the action and camera; change nothing else. Generate at least two takes per scene: retrying is the normal cost of this workflow, not a sign you’re doing it wrong.

Step 4: Chain scenes with first and last frames

The trick that makes five clips feel like one film is first/last frame chaining: the final frame of clip A becomes the starting frame of clip B, so the join is a continuous image rather than a hard cut. The getimg.ai guide explains it plainly: the video begins from the first image and transitions into the last one.

Veo 3.1 supports first- and last-frame transitions directly (Google’s Veo page describes it as creating “smooth, artful, and epic transitions between images”) and Kling carries start/end frame support through its current models. Luma Ray3.2 goes further, with up to 16 keyframes inside one 20-second clip plus start/end keyframe linking for the seams between generations.

To use it: export the last frame of scene 1 and load it as scene 2’s first frame. Keep the endpoint frames compatible (same character position, same lighting family) or the model will morph one into the other in ways you didn’t ask for. Want a camera move between scenes instead of continuity? Make the endpoint frames deliberately different and let the transition do the storytelling.

Step 5: Add dialogue and lip sync

You have three tiers, in ascending order of control. Veo 3.1 generates dialogue, effects, and ambience natively with the clip (no extra step) though Google’s documentation notes that natural, consistent spoken audio, especially in short segments, is still being refined. Kling 4.0 does lip-synced dialogue in one pass, and you can bind a voice to a character with a 5-30-second clean recording per the official guide, keeping face and voice consistent across scenes.

The third tier is post-hoc lip sync, applied to a finished clip. Kling’s Lip Sync feature animates an existing video from an audio file or typed text, and standalone tools do the same: Hedra’s Character-3 for audio-driven talking heads, sync.so for precise mouth re-timing, and open-source options like LatentSync and MuseTalk if you run your own stack. Post-hoc sync is also your fix when a take has the right acting but the wrong mouth.

One warning from the docs rather than the hype: lip sync degrades with fast, overlapping, or very short lines. Write for the technology: one speaker, complete sentences, moderate pace.

Cut the five clips into one film

Any editor works; CapCut and DaVinci Resolve are both free. Cut on motion so the eye doesn’t inspect the join, duck the music under dialogue, and lay one grade over all five clips: a shared look hides small drift better than any prompt. Sound design is the cheapest repair: a wind bed and one motivated sound effect will sell a transition that visuals alone can’t.

This is also where you fix instead of regenerate. A slightly off face for six frames? Trim it. A hand that melts? Cut away to the storm. Editing is the last consistency tool, and it never charges credits.

What still goes wrong (the honest list)

Identity drift (the character changing appearance between scenes) is reduced, not solved. Even with references, creators on Google Flow and similar platforms still report characters shifting across long projects: a subtly wider face, lighter eyes, hair that restyles itself. Treat every generation as a draft.

The documented weak spots cluster. Costume changes break identity because models anchor partly on clothing, if the story needs an outfit change, make it between scenes and supply a new reference for the new outfit. Side profiles and extreme angles are the second failure case, since frontal faces anchor best; that’s why the reference sheet includes a profile. Fast dialogue is the third, where lip sync gets mushy.

Two workflow realities. First, these tools churn monthly: Google retired the Veo 3.1 preview endpoint in the Gemini API on October 22, 2026, and Kling 4.0 hadn’t reached the API as of late October: re-check specs before budgeting. Second, the retry tax is real: at Veo’s standard API rate of $0.40 per second, one 8-second scene costs $3.20, so five scenes at two takes each runs about $32, arithmetic from the listed rate, not a quoted package price.

And if the film becomes commercial work: commercial rights come from each platform’s terms, not the model. Runway’s help center says your outputs are yours to use without non-commercial restrictions, while Kling ties commercial rights to paid memberships per its user policy. Ownership of purely AI-generated work is legally unsettled in several jurisdictions, so this is where real legal advice is worth the fee.

Where to start

If you want dialogue and simplicity: Veo 3.1 in Google’s Flow with a Google AI Pro plan. If you want maximum control and don’t mind a steeper interface: Kling 4.0, starting at $10/month. If you’re already an image-first creator: Runway. On zero budget: Vidu’s free credits or Pika Character Studio.

Whichever you pick: write one identity block, generate one reference sheet, and animate a single 8-second scene from it three times. Once the same face comes back three times, the other four scenes are just repetition with better stakes.

FAQ

What is reference-to-video and how does it keep a character consistent?

Reference-to-video is a generation mode where you upload still images of a character, and the model animates scenes around those images instead of inventing a new person each run. The images anchor facial structure, hair, and clothing, so the same character appears in every clip. Kling calls it Omni Reference, Google calls it Ingredients, Runway calls it References.

What’s the best AI video tool for one consistent character?

Kling 4.0 is the most capable overall as of November 2026: up to 10 reference images, 10 keyframes, 30-second clips, and native lip-synced dialogue. Veo 3.1 is best for dialogue-first stories, and Runway Gen-4 does the most from the least: a single reference image. Choose by story type, not by leaderboard.

How do I make a five-scene story with the same AI character?

Write a fixed identity block, generate a multi-angle character reference sheet with an image model like Nano Banana Pro, then generate each scene with reference-to-video using the same sheet and wording. Chain clips by feeding each scene’s last frame into the next scene’s first frame, then cut the five clips together with music and one shared grade.

How many reference images do these tools support?

Kling 4.0 accepts up to 10 images plus 5 videos (15 combined assets), Vidu takes 1-7 reference images, Hailuo’s H3 Max supports up to 9 via API, and Veo 3.1 takes up to 3 “ingredients.” Runway Gen-4 needs only one image for character consistency, and Luma Ray3.2 pairs character references with up to 16 keyframes per clip.

How much does a five-scene AI video cost?

With Veo 3.1’s standard API rate ($0.40 per second, audio included), five 8-second scenes cost $16 at one take each, roughly $32 at two takes. Kling’s Pro plan ($37/month, 3,000 credits) covers about 250 seconds at 1080p with audio. Runway charges 12 credits per second; its $12 Standard plan includes 625 monthly credits. Expect to spend real credits on retries.

How do I make my AI character talk with lip sync?

Three ways: generate the clip with dialogue natively (Veo 3.1 or Kling 4.0 do this in one pass), use Kling’s Lip Sync feature to animate an existing clip from audio or text, or use a standalone tool like Hedra’s Character-3 or sync.so. Keep lines short, one speaker per clip, at a moderate pace.

Why does my character still change between scenes?

Identity drift persists because models anchor identity partly on frontal facial features and clothing. Costume changes, side profiles, extreme camera angles, and fast motion are the documented failure cases. The fixes: keep the outfit constant, include a profile view in your reference sheet, reuse the exact identity wording, and generate multiple takes per scene.

How do I make smooth transitions between AI video scenes?

Use first/last frame chaining: export the final frame of one clip and load it as the first frame of the next, so the join is continuous. Veo 3.1 supports first- and last-frame transitions, Kling supports start/end frames, and Luma Ray3.2 offers up to 16 keyframes plus start/end keyframe linking. Keep endpoint frames’ lighting and position compatible.

Can I sell videos made with these AI tools?

Commercial rights come from each platform’s terms, not the model. Runway’s help center states your outputs are yours to use without non-commercial restrictions, while Kling ties commercial rights to paid memberships: free-tier output generally isn’t cleared for commercial use. Read the current terms before monetizing, and get legal advice for anything high-stakes, since AI-generated work’s copyright status varies.

Which AI video tool is best for beginners?

Start with Veo 3.1 in Google’s Flow: the interface is the friendliest of the bunch, 8-second clips with audio need no editing chops, and a Google AI Pro subscription bundles the credits. Build your reference sheet in the same Google ecosystem with Nano Banana Pro. Move to Kling 4.0 once you want longer, more controlled shots.

Sources

• Kling 4.0 model release page: Kling AI
• Kling AI Officially Launches Next-Generation Kling 4.0: Kuaishou via GlobeNewswire
• Kling VIDEO 3.0 Omni Model User Guide: Kling AI
• Kling AI Lip Sync: Kling AI
• Kling AI membership plans: Kling AI
• Veo 3.1 Ingredients to Video: Google blog
• Veo model page: Google DeepMind
• Veo documentation: Google AI Studio
• Ultimate prompting guide for Veo 3.1. Google Cloud
• Veo 3.1 listing with API pricing: OpenRouter
• Introducing Runway Gen-4. Runway
• Runway pricing page: Runway
• Creating with Gen-4.5: Runway Help Center
• Output ownership and commercial use. Runway Help Center
• Introducing Ray3.2. Luma AI
• Luma pricing. Luma AI
• Reference to Video: Vidu
• S2V-01 Subject Reference release: MiniMax
• Character Studio. Pika
• Guide to first and last frame control: getimg.ai
• Character consistency in AI video generation: Rendar
• AI lip sync: Hedra