AI reference to video · MiniMax H3 · Wan 3.0 · Seedance 2.5 · Seedance 2.0 · Seedance 2.0 Fast · Seedance 2.0 Mini · Kling 4.0 Flash · Kling 4.0 · Kling O1 · Kling 3.0 Omni · Kling 3.0 · Kling 2.6

Reference to Video AI Generator: Same Character, Any Scene

Upload 1–5 reference images, describe a new scene, and get a 2–15 second MP4. More images, clips and sound depend on the model.

1 · Model12 live
2 · Referencesimage, video, audio
Images(0/10)Face, product or place
JPG/PNG/WebP · ≤10MB
Videos(0/5)Motion to follow
MP4/MOV · 2–15s · ≤50MB
Audio(0/5)Voice to match
MP3/WAV · 2–15s

Need 1+ image or video · videos 15s total, billed per second like output

3 · Prompt

Name uploads as labelled: Image 1, Video 1, Audio 1…

0 / 5000

4 · Settings
Shape
5s
Quality

Native sound included

600 credits for this run

Describe what you want to make.

Uploads deleted after 24 h · No watermark · Credits back if a run fails

Meadow editorial lock

Wan 3.0 · 16:9 · 5s · 720p

Use Image 1 as the performer. She is reclining across a lime-green inflatable modular sofa in the middle of a wildflower meadow, wearing a white hooded coverall with a ruffled panel down the front, white trousers and pale platform shoes, with a small pale-blue object tucked against her hip. She looks straight down the lens and does not move, and only her head turns slightly as a large orange butterfly settles on her outstretched arm. More butterflies drift through the flowers around the sofa. Keep her face, hair and the whole outfit exactly as they appear in Image 1. Fashion editorial film, flat frontal composition with the sofa centred, soft overcast light with no visible shadows, a heavily saturated grade that pushes the grass and the sky toward the same green, an almost imperceptible slow push in, shallow depth of field on the foreground flowers, fine grain. Wind through grass and one distant bird, no music and nobody speaking.

Click any sample clip to preview2 / 6
Subject Lock -> New Scene

What Can Our Reference to Video Examples Keep Consistent?

Inspect each reference-to-video demo beside its source subject plates, slot-numbered prompt, model settings, credit cost, and native-sound MP4. Click Use this prompt to load the same setup above.

Text PromptMiniMax H3 · 9:16 · 15s · 720p · 525 credits

Giant croissant UGC

Create a UGC video ad featuring [image1], presenting product [image2] 15-second vertical 9:16 authentic UGC bakery sharing video, shot on a phone by a friend, bright natural daylight, casual handheld micro-shake, photorealistic, real skin texture, exact English lip-sync, no subtitles, no text overlays, no watermark. Use the reference person: stylish young Black woman with dark skin, braided hair styled into two low buns with long face-framing braids, wearing a wide black hat, black sleeveless fitted top, white mini pleated skirt, white collar detail, black-and-white belt, soft confident expression. No handbag. Background: cozy modern croissant bakery storefront, warm beige facade, big glass window full of pastries, outdoor wooden tables and chairs, soft golden afternoon light, and a giant croissant sign on top of the shop. Hero food: an oversized giant croissant, almost half her body size, golden brown glossy crust, flaky layered ridges, buttery texture, crisp edges, warm bakery shine. Shot 1: medium full-body outside the bakery. She holds the giant croissant with both arms, laughs at how huge it is, faces camera with mouth visible and says: "Okay, this is officially the biggest croissant I've ever held." Shot 2: close-up of the croissant. She tilts it toward the camera to show the glossy crust, flaky layers, and huge size, then says: "Look at these layers. It's literally bigger than my bag." Shot 3: she takes one real bite from the edge, crunch sound, crumbs fall naturally. She lowers the croissant so her mouth is visible, reacts with widened eyes and says: "Wait. It's actually so buttery and crispy." Shot 4: she hugs the giant croissant playfully, standing in front of the bakery window, and says: "I thought this was just for photos, but no, I'm eating this." Final shot: full-body bakery exterior, giant croissant visible in her arms, soft smile to camera: "If you see this place, get the giant one." Natural street ambience, bakery cafe sounds, crispy bite, light crumbs, live on-camera speech, exact lip-sync. The giant croissant stays flaky, golden, and clearly visible throughout.

Reference inputs
Character reference: creator in wide black hat, black top, and white pleated skirtPerson ref
Object reference: oversized golden croissant held by bakerProduct ref
Compare engines ↓
Sample Video

Editorial demos generated on Wan 3.0 and MiniMax H3 showing 1–10 input references, locked appearance, and native audio.

Live Multi-Ref Engines

Which Engines Run Inside This Reference to Video AI Generator?

Our reference to video ai generator pairs Wan 3.0 (2–15s, explicit Image 1–10 calling) and MiniMax H3 (4–15s, flexible portrait/wide tolerances) on one upload rack. Need clean character sheets first? Build plates in our 4K image tools.

R2V Engine Specifications
10 Specs
Side-by-side comparison of Wan 3.0 and MiniMax H3 for reference to video generation
SpecWan 3.0 (default)MiniMax H3
Upload slots (our studio)1–10 photos1–10 photos
Upstream slot ceilingUp to 10 imagesUp to 9 images
Single upload size limit≤ 20 MB per photo≤ 20 MB here (JPG, PNG, WebP, HEIC)
Prompt naming ruleCall Image 1, Image 2… in upload orderState who does what clearly per role
Clip length2–15s (default 5s)4–15s (default 8s)
Aspect options16:9, 9:16, 1:1, 4:3, 3:416:9, 9:16, 1:1, 4:3, 3:4, 21:9
720p credits / s120 cr/s35 cr/s
1080p credits / s240 cr/s55 cr/s
Prompt limit5,000 chars7,000 chars
Best forFast 2–15s multi-subject clips & products4–15s spoken UGC, 21:9 cinema & deep scripts
← Swipe horizontally to read full table →

Summary: start on Wan 3.0 (120 cr/s at 720p) for rapid reference to video blocking, or pick MiniMax H3 for 4–15s dialogue scenes.

How Subject Locking Works

How Does AI Reference to Video Lock Faces, Pets, and Logos?

Instead of freezing your upload as frame one, ai reference to video extracts identity features from 1–10 reference photos so you can move the character into a completely new angle, outfit, or setting.

Reference-to-Video vs. Image & Text Modes

Alibaba’s Wan 3.0 video generation guide notes reference images render new scenes "with consistent appearance", while MiniMax’s video generation API treats them separately from start/end frames. Use Text to Video to invent worlds, Image to Video to animate frame one, and this reference to video generator when a fixed face or SKU must enter fresh compositions.

Re-Composes AngleKeeps Face & OutfitNative Sound
Identity LockGenerate MP4

Use Image 1 as the performer... Keep her face, hair and the whole outfit unchanged.

Which Reference Angles Prevent Face Drift

One headshot isn’t enough if your character turns sideways: fal’s reference-to-video guide shows faces drift without a profile angle and warns against cramming unrelated styles ("reference overload"). Keep each slot single-subject on plain light, and fill only the 1–10 slots your shot truly needs.

1 Subject / SlotFront + ProfileClean Lighting
Multi-Slot LockGenerate MP4

Maintain the exact identities, hairstyles, outfits and equipment of both characters.

Name Image Slots Explicitly & Know Current Bounds

Alibaba’s Wan 3.0 guide names slots directly ("Using the lip gloss product in Image 1…"), while HeyGen’s R2V guide stresses letting photos hold looks while words steer action—never write "blonde woman" over a brunette upload.

Name Image 1 / Image 2Action Over LookAudio Included
Prompt FormulaGenerate MP4

Subject tag (Image 1/2) + Action + Setting + Camera + Lock cue + Sound.

Parameters & Reference Guides

3 Tables · 1 Note
3 Video Modes Compared
3 Modes
When to pick Text to Video, Image to Video, or Reference to Video
CriterionText to VideoImage to VideoReference to Video
InputsPrompt text only1 start frame (+ optional last)1–10 reference images + prompt
Opening compositionModel invents everythingLocked to frame one photoFreely recomposed around subjects
Best when you wantBrand-new scenes fastSubtle motion inside one photoOne person, pet or product across shots
← Swipe horizontally to read full table →

Summary: pick a consistent character video ai workflow whenever framing changes but your hero cannot.

5-Slot Allocation Guide
5 Slots
Recommended role allocation for Image 1 to Image 5 plus photo checklist
SlotWhat to uploadTip for clean lock
Image 1 (Required)Front hero view of person, pet or SKUUncovered face or logo, sharp focus
Image 2 (Optional)Side / 3-quarter angle OR co-star / propStops identity morphing during head turns
Image 3 (Optional)Held product, co-subject or outfit detailName explicitly as Image 3 in prompt
Image 4 (Optional)Target room, street or studio backgroundMatches color temperature of subjects
Image 5 (Optional)Extra object or lighting mood plateSkip unused slots to avoid overload
← Swipe horizontally to read full table →

Summary: evenly lit, uncropped photos inside aspect 0.4–2.5 keep every subject separated cleanly.

Weak vs. Strong Prompts
3 Examples
Weak vs strong reference-to-video prompts for single subject, UGC product, and two-character shots
Scene TypeWeak (causes drift)Strong (slot-anchored)
One character"Woman walking in neon rain""Image 1 walks past neon cafe windows, medium tracking shot, rain foley"
Host + SKU"Girl holding skincare bottle""Person in Image 1 raises serum bottle Image 2 toward camera, soft morning kitchen"
Two characters"Two astronauts talking on deck""Navigator Image 1 greets operative Image 2 inside hangar; preserve outfits, low two-shot"
← Swipe horizontally to read full table →

Summary: assign every upload (Image 1, Image 2…) one clear verb so multiple images to video never swap faces.

Current bounds & simple fixes

Five real limits apply today: max clip duration is 15s (stitch 5–15s shots in an editor for longer reels); 180-degree turns need Image 2 as a profile photo; multiple faces near each other can blend unless separated by slot number; tiny label typography may soften in motion (pair slow push-ins with clean studio crops); and sound is synthesized natively rather than cloned from an uploaded voice.

3-Step Studio Flow

How Do You Make a Reference to Video Clip in 3 Clear Steps?

Upload 1–10 clear photos, reference Image 1, Image 2 in your prompt, check exact credits on the Generate button, and download your MP4.

  1. STEP 01Step 01 · Refs
    Reference to video AI generator top panel showing Wan 3.0 selector and Image 1 upload slot

    1. Upload 1 to 5 Reference Photos

    Pick Wan 3.0 or MiniMax H3, drop JPG, PNG, WebP, or HEIC stills up to 20 MB each, and set upload order carefully—Image 1 holds your main hero.

  2. STEP 02Step 02 · Prompt
    Prompt box referencing Image 1 with shape, length slider, and 720p or 1080p controls below

    2. Name Slots, Pick Shape, Length & Quality

    Write what Image 1 (and Image 2…) should do, pick 16:9, 9:16, 1:1, 4:3, 3:4, or 21:9 (H3), choose 2–15s length and 720p or 1080p, and read live credits upfront.

  3. STEP 03Step 03 · Export
    Preview player displaying finished 9:16 reference-to-video UGC clip with sound toggle

    3. Preview Sound & Export Clean MP4

    Press Generate, review identity lock with native sound on or muted, then save your zero-watermark H.264 MP4—stored in My creations for 30 days.

Full Upload & Output Specification Table

Upload & Export Rules
9 Specs
Technical rules, limits, and retention behavior for reference to video generation
SettingRule in Our Studio
Reference slots1–10 photos per clip
Formats & sizeJPG, PNG, WebP (plus HEIC on H3) · ≤ 20 MB per file
Aspect toleranceKeep each photo inside 0.4–2.5 W/H ratio
Slot namingAddress as Image 1, Image 2… in exact upload order
Prompt limit5,000 chars (Wan 3.0) · 7,000 chars (MiniMax H3)
Length & shape2–15s (slider) · 16:9, 9:16, 1:1, 4:3, 3:4 (+ 21:9 on H3)
Resolution720p or 1080p clean MP4
Native audioOn-screen dialogue, foley & ambience muxed automatically
Retention & refundsUploads auto-delete after 24h · MP4s stay 30 days · errors auto-refund
← Swipe horizontally to read full table →

Summary: 1–10 photos (≤ 20 MB each) render 2–15s MP4s with native sound and automatic upload purging.

Consistent Storytelling

Who Uses a Consistent Character Video AI Generator?

Creators, e-commerce brands, and episodic storytellers use a consistent character video ai workflow to place one verified persona, pet, or SKU into unlimited settings without reshoots.

UGC & EcommerceMiniMax H3 · 9:16 · 15s

Combine One Creator & One SKU in Any Location

Drop your spokesperson in Image 1 and packaging in Image 2 to shoot storefront, kitchen, or unboxing variations in minutes.

Starter Recipe

Create a UGC clip featuring Image 1 holding bakery pastry Image 2 outside warm cafe, live reaction speech.

Inspect Every Live Reference-to-Video Output

Click any clip above in the studio to test multi image to video prompts right away.

Giant Croissant UGC
H3 · 9:16 · 15s
Meadow Editorial Lock
Wan 3.0 · 5s
Sci-Fi Trailer Cuts
H3 · 16:9 · 15s
Two-Subject Exchange
Wan 3.0 · 5s
Gothic Drama Teaser
H3 · 9:16 · 15s
Desert Train Heist
Wan 3.0 · 5s
Transparent Credit Math

How Much Does a Reference to Video Generator Cost Per Clip?

Every parameter change updates credits live on the Generate button: our default Wan 3.0 5s 720p clip costs 600 credits. Adding 1 or 10 photos costs the exact same rate—billing depends only on engine, length, and quality. See Pricing.

Credits by Engine & Length
5 Credits
Credit cost by clip duration, model, and resolution in our reference to video generator
Clip LengthWan 3.0 (720p)MiniMax H3 (720p)1080p upgrade
2s quick test240 cr— (min 4s)From 240 cr/s
5s standard clip600 cr (default)175 cr275 cr (H3)
8s social hook960 cr280 cr1,920 cr (Wan 3.0)
10s story beat1,200 cr350 cr55 cr/s (H3)
15s UGC ad1800 cr525 cr825 cr (H3)
← Swipe horizontally to read full table →

Summary: 240–1800 credits at 720p regardless of whether you upload 1 or 10 reference photos.

Packs, Plans & Per-Second Cost
6 Tiers
Credit packs and monthly plans rated by Wan 3.0 5s 720p clips (600 credits each)
TierPriceCredits~5s Wan Clips ($/s)
Small pack$9.901,500~2 ($0.79/s)
Medium pack$24.904,000~6 ($0.75/s)
Large pack$49.909,000~15 ($0.67/s)
Starter plan$14.90/mo3,000~5 ($0.60/s)
Pro plan$24.90/mo9,000~15 ($0.33/s)
Premium plan$99/mo40,000~66 ($0.30/s)
← Swipe horizontally to read full table →

Packs never expire; credits also unlock AI video templates and 4K image tools.

Stretch Credits on Multi-Image Shoots

  1. 1Test blocking in Wan 3.0 · 2s or 5s at 720p (240 or 600 cr) before paying for 15s 1080p MiniMax H3.
  2. 2Lock front + profile refs once in our reference to video generator so you spend zero retries fixing mid-turn drift across scenes.

Unused first packs qualify within 7 days under our refund policy; failed or upstream-blocked runs refund automatically.

Privacy, Models & Disclosure

Why Do Creators Trust This Multi Image to Video Workspace?

Reference uploads run directly over Alibaba Cloud Model Studio and the MiniMax API in slot order with strict automatic deletion—see our privacy policy and terms.

01

24h Upload Purge · Never Trained

Reference photos auto-delete after 24h and finished MP4s expire after 30 days. Only upload adults (18+) with explicit permission.

02

Clean Upstream Forwarding

Your Image 1–10 stack and exact prompt reach Wan 3.0 or MiniMax H3 untouched; failed runs refund every credit automatically.

03

Zero-Watermark 720p / 1080p MP4

Finished clips store without watermarks or extra compression, ready for client ads, social campaigns, and multi-shot edits.

04

Built-In Platform Labeling Readiness

Add AI labels for realistic footage on TikTok and YouTube—YouTube notes disclosure never limits reach.

Reference to Video FAQ

Got Questions About Reference to Video AI?

1.How does reference to video AI keep a character consistent across shots?

Each render reads appearance cues—facial structure, outfit seams, fur markings, or logo geometry—from your 1–10 reference photos and rebuilds the requested new angle, lighting, and movement around those subjects.

2.How long does it take to render from reference photos?

Queue speed depends on duration, resolution, and server load. Results wait in My creations for 30 days.

3.Can I upload a photo of a real person for AI reference to video?

Only when you own the photograph or hold consent from every depicted adult (18+ strictly required). Non-consensual impersonation violates our terms.

4.Do multiple images cost more credits than a single reference image?

No. Uploading 10 images costs the exact same credits as uploading 1—billing depends purely on engine (Wan 3.0 vs MiniMax H3), seconds, and 720p vs 1080p.

5.How many reference photos can I upload here?

Our studio supports 1–10 reference photos per job (20 MB each). In practice, 2–3 focused photos per subject outperform 5 crowded ones.

6.How is this different from image to video or Vidu / HeyGen?

Unlike frame-one animation, reference to video freely changes camera angles around your subject. Vidu accepts 1–7 refs on its own models and HeyGen’s R2V runs on MiniMax H3; here you swap Wan 3.0 and MiniMax H3 side-by-side with upfront per-second credits and zero watermarks.

7.Why did my character’s face change mid-turn or swap outfits?

Usually the character turned further than the uploaded angle showed, two similar photos overloaded a slot, or the prompt contradicted Image 1 (e.g., naming another hair color). Add a clean side view in Image 2 and delete visual adjectives that conflict with the photo.

8.How long are uploaded reference images stored?

Uploaded reference photos auto-delete after 24 hours and finished MP4 clips delete after 30 days. Nothing is used to train models or shown publicly.

9.Can I use these multi image to video clips for ads and commercial work?

Yes—whenever you own or license every uploaded person, logo, or product photo. Add platform AI disclosure labels for photorealistic campaigns on TikTok and YouTube.
Ready to Direct

Ready to Put Your Character Into Its Next Scene?

Drop 1 to 10 clear photos above, call them Image 1, Image 2 in your ai reference to video prompt, and check exact upfront credits before rendering.