Most AI video models give you one clip. You describe a scene, they generate it. If you want a story - establishing shot, mid shot, close-up - you run three separate generations, then manually assemble them in CapCut, hoping the character still looks like the same person from shot to shot.
Wan 2.6 takes a different approach. Structure your prompt with temporal markers - Shot 1, Shot 2, Shot 3 - and it generates the full sequence in a single pass, maintaining character consistency, visual continuity, and scene logic across the cuts.
This guide covers all three of Wan 2.6's generation modes, the exact prompt structures that activate each capability, and where it fits against other video models on Cliprise.
What Wan 2.6 Is
Wan 2.6 is Alibaba's video generation model, built on a 14B parameter Mixture-of-Experts Diffusion Transformer architecture. It was released December 2025 and distinguished itself from contemporaries with three specific capabilities that no other model on Cliprise replicates exactly:
Multi-shot narrative generation. A single prompt can describe multiple distinct scenes with transitions, and Wan 2.6 generates a coherent video sequence rather than one continuous clip. The model maintains character consistency, lighting direction, and visual style across each shot.
Reference-based generation (R2V). Upload a 2–30 second reference video of a person, and Wan 2.6 extracts their appearance and voice characteristics. Then generate new scenes featuring that same character - identity stays consistent across each new clip you produce.
Native audio-video generation. Audio and video are generated together in a single pass, not assembled in post. This includes phoneme-accurate lip sync, facial micro-expressions and jaw movements aligned to speech, ambient sound, and music where appropriate.
Technical specifications:
- Resolution: up to 1080p
- Duration: up to 15 seconds (T2V, I2V); up to 10 seconds (R2V)
- Aspect ratios: 16:9, 9:16, 1:1
- Languages: English and Chinese prompts
- Architecture: 14B parameter MoE Diffusion Transformer
The Three Generation Modes
1. Text-to-Video (T2V) - Including Multi-Shot
Standard T2V produces a single clip from a text description. This works well for one specific scene. For narrative content, the multi-shot structure is significantly more powerful.
Single-shot prompt (standard use):
A barista preparing espresso in a morning café,
overhead angle looking down at the cup as coffee extracts,
warm amber light, rising steam, slow deliberate motion,
professional food/beverage cinematography
Multi-shot prompt (what makes Wan 2.6 distinctive):
Overall visual world: Morning café, warm amber tones,
soft natural light from large windows, film grain texture.
Shot 1 [0-4s]: Wide shot of empty café interior at dawn,
chairs still upturned on tables, warm light entering from left,
one barista moving through background, peaceful atmosphere.
Shot 2 [4-9s]: Medium shot, barista at the espresso machine,
hands working deliberately, steam rising from portafilter,
face in partial profile, focused expression.
Shot 3 [9-14s]: Close-up overhead on espresso cup as coffee extracts,
rich golden crema forming, slow deliberate pour,
shallow depth of field.
The model reads the shot structure and generates all three in a single pass - visual style, lighting direction, and the barista character remain consistent across cuts. You get a 14-second short film rather than a single isolated moment.
Shot marker format:
Shot [number] [[start time]s-[end time]s]: [camera description], [action description], [mood/quality]
Time markers should add up to your target duration (maximum 15 seconds). Three 5-second shots work cleanly. Two shots of 7 seconds each also works. The model interprets the proportion rather than exact frames.
2. Image-to-Video (I2V) - Animate Any Image
Upload an existing photo or AI-generated image as the starting frame and describe the motion. Wan 2.6 produces a clip that begins from that exact image and animates it.
What this is useful for:
- Animating a Flux 2 or Midjourney generated image into a clip
- Adding motion to product flat lays
- Turning still illustrations into moving scenes
- Creating B-roll from existing photography
Prompt structure for I2V:
The image defines the visual starting point. The prompt describes motion and what changes - not what is in the image, as the model can already see that.
Good I2V prompts focus on:
- Camera movement (slow pull-back, subtle push-in, slight pan left)
- Environmental elements in motion (leaves moving, light shifting, water surface)
- Subject movement if the image contains a person (slow turn, breathing, subtle gesture)
Examples:
For a product shot:
Camera very slowly orbits clockwise around the product,
gentle parallax between foreground and background,
soft light catches the surface texture,
5 seconds, smooth and premium
For a portrait:
Subtle natural breathing movement, slight hair movement from ambient air,
soft environmental light shifts slightly warmer,
natural blink at 2 seconds, serene and still
For a landscape:
Slow push into the scene, clouds moving gently across the sky,
light quality shifts subtly as clouds pass,
foreground elements have gentle natural movement
I2V multi-shot: You can combine I2V with the shot structure. Upload an image that defines the visual world, then write shot markers that animate from that starting frame into different compositions.
3. Reference-to-Video (R2V) - Consistent Character Across Clips
R2V is the most distinctive Wan 2.6 capability and the one with the fewest direct equivalents on other models.
How it works:
- Record a 2–30 second reference video. Clear face, natural movement, speaking if you want voice characteristics included.
- Upload as reference in Wan 2.6 R2V mode on Cliprise.
- The model extracts appearance and movement characteristics from the reference.
- Write a prompt describing the new scene.
- The model generates the new scene featuring the same character.
What consistent means here: The character's face, build, hair, skin tone, and movement style carry across from the reference into every new clip you generate from it. The person does not drift into someone who looks vaguely similar - they look like themselves.
Multi-character R2V: Upload up to three separate reference videos. Each reference is assigned a tag in the prompt (Reference 1, Reference 2). The model places each character into the scene with their respective identity.
What works well as reference:
- Good lighting, face clearly visible
- Natural conversational movement - not stiff or posed
- At least 5 seconds to give the model enough signal
What produces less reliable results:
- Reference video with heavy backlight or very low light
- Face partially occluded
- Very short reference (under 3 seconds)
- Heavy makeup or accessories that significantly alter base appearance
R2V prompt structure:
[Character name from reference, e.g. "the presenter from Reference 1"]
[New scene description]
[Environment]
[Camera and framing]
Working example:
Reference 1 character - the woman from the reference video -
standing at a bright kitchen counter, morning light from window on her left,
explaining something naturally to camera, warm and confident tone,
medium shot, slight depth of field behind her,
10 seconds
R2V for recurring brand content: Create a brand spokesperson video library from a single reference recording. Generate multiple scenes - product explanation, FAQ, onboarding step - all featuring the same face, without re-recording each time.
Prompting for Quality Results
What Wan 2.6 Responds To
The model responds well to cinematography language:
- Camera direction: wide establishing shot, medium two-shot, close-up, overhead, Dutch angle, POV
- Camera movement: slow push-in, pull-back, tracking left, subtle orbit, handheld energy, static locked-off
- Lighting: soft diffused daylight, dramatic side lighting, warm practical lights, overcast exterior
- Quality descriptors: film grain, shallow depth of field, professional cinematography, high production value
Consistency Across Shots
When writing multi-shot prompts, the "overall visual world" paragraph at the top matters. Put color palette, light quality, era, and visual style information there. Each shot description then references the specific moment - not re-describes the world. The model uses the world description as context for every shot.
If character appearance is important across shots, mention consistent character details in the world description:
Overall visual world: Contemporary apartment interior, warm evening light.
Consistent character: woman, early 30s, dark shoulder-length hair,
grey linen blazer, natural confident presence.
Shot 1 [0-5s]: ...
Shot 2 [5-10s]: ...
Aspect Ratio and Platform Format
- 9:16 (vertical) for Reels, TikTok, YouTube Shorts
- 16:9 (horizontal) for YouTube standard, website embedding
- 1:1 (square) for Instagram feed posts, some social platforms
Specify the target format in the prompt - "vertical 9:16 composition, subject centered" - and Wan 2.6 will adjust framing accordingly.
When to Use Wan 2.6 vs Other Models
| Need | Best model | Reason |
|---|---|---|
| Multi-shot narrative in one generation | Wan 2.6 | Native shot planning |
| Same character across multiple clips | Wan 2.6 (R2V) | Reference-based consistency |
| Highest single-shot visual quality | Kling 3.0 | 4K native, stronger photorealism ceiling |
| 4K resolution output | Kling 3.0 | Wan 2.6 caps at 1080p |
| Physics simulation, water, weather | Veo 3.1 | Environmental physics accuracy |
| Music-synchronized video | Seedance 2.0 | @Audio tag for music-responsive generation |
| Fast short social clips | Kling 2.5 Turbo | Speed optimized |
| Stylized / painterly aesthetic | Hailuo 02 | Distinct visual character |
The practical workflow many creators use: Wan 2.6 for multi-shot narrative foundations and character-consistent content, Kling 3.0 for hero shots that need maximum visual quality, assembled in CapCut.
See Wan 2.6 vs Kling 2.6: Chinese AI Video Models →
Production Workflow: 15-Second Social Narrative
Complete workflow for a short-form narrative clip:
-
Write your shot list before touching the model. Three shots, roughly 5 seconds each. What happens in each shot, what camera angle, what connects them. A sentence per shot on paper is faster than iterating in generation.
-
Structure the prompt with the world description at top, then shot markers. Keep shot descriptions concise - the model works better from clean descriptions than long paragraphs per shot.
-
Generate in T2V multi-shot mode. Review the transitions: does character appearance stay consistent shot to shot? Does the visual style match across the cuts?
-
If a specific shot is weak: Use I2V on the frame from the prior shot as a starting point for the next one. This gives you tighter visual continuity where multi-shot drifts.
-
Audio: Wan 2.6 may generate native audio based on the scene. If you want scripted narration over it instead, mute the generated audio in CapCut and replace with ElevenLabs TTS narration.
-
Export: 9:16 for Reels and TikTok, 16:9 for YouTube. Add captions in CapCut using auto-captions or the transcript from ElevenLabs Speech-to-Text.
Note
Wan 2.6 is available on Cliprise alongside Kling 3.0, Veo 3.1, Seedance 2.0, and 40+ other video models. Try Cliprise Free →
Related Articles
Wan model comparisons:
Video generation guides:
- Alibaba's video ecosystem now also includes HappyHorse 1.0 on Cliprise, which adds another practical option for short-form generation, image-to-video, reference-driven clips, and marketing video workflows. Wan remains important for broader AI video experimentation and multi-shot workflows, while HappyHorse is especially worth testing when the brief starts from a product image, app mockup, or reference subject. The dedicated model landing page is HappyHorse 1.0.
- AI Video Generation 2026: 22+ Models, Workflows →
- Image-to-Video Workflow: Complete Guide →
- Best AI Video Models on Cliprise 2026 →
Prompting:
Video model comparisons:
Chaining models:
Models on Cliprise:
Published: March 19, 2026.
