WAN 2.6 Reference-to-Video
Wan 2.6 Reference-to-Video learns style, motion and framing from 1–2 reference clips. 5–10s at 720p or 1080p, native audio, 50/80 credits per second.
Wan 2.6 Reference-to-Video generates a new shot from one or two reference clips. It learns the style, motion, framing and pacing of the references, then applies them to a new scene described in the prompt. Output runs 5 or 10 seconds at 1280×720, 720×1280, 1920×1080 or 1080×1920. Pricing is 50 credits per second at 720p and 80 at 1080p — and unlike Wan 3.0, only the output duration is billed, not the reference length.
- References: 1–2 videos
- Duration: 5s / 10s
- Price: 50–80 credits/s
- Audio: Native, included
How to use WAN 2.6 Reference-to-Video
Reference clips set the grammar; the prompt sets the new sentence. Keep both simple.
- Choose one or two reference clips Pick footage with clear motion, framing and pacing. Two clips should agree on style, or the model averages them into something generic.
- Describe only the new scenario The references already carry appearance and motion language. Spend the prompt on the new subject, setting and action.
- Set duration and shot type Five seconds for a single beat, ten for a short sequence. Multi-shot is useful when the new scene needs a cut.
- Choose size and resolution 720p costs 50 credits per second, 1080p costs 80. Draft at 720p and deliver at 1080p only when the detail matters.
- Add audio only if it is part of the brief Native sound is included; uploading a track is optional. Leave it out when you will add music in the edit.
Settings reference
WAN 2.6 Reference-to-Video takes one or two reference clips and applies their style, motion and framing to a new scene. This endpoint does not accept reference images.
| Setting | Value | Notes |
|---|---|---|
| Reference videos | 1–2 clips, required | Up to 100 MB per upload. The model learns style, motion, framing and pacing. |
| Prompt | Required | Describe the new scenario for the subject from the references. |
| Audio upload | Optional, up to 15 MB | Supply a track when the result must follow specific audio. |
| Duration | 5s / 10s | Default is 5 seconds. |
| Size | 1280×720 / 720×1280 / 1920×1080 / 1080×1920 | Landscape and portrait at 720p and 1080p. Default is 1280×720. |
| Shot type | Single / Multi | Multi-shot lets the output cut between beats. |
| Prompt expansion | Optional toggle | Expands sparse prompts into fuller scene descriptions. |
| Seed | Optional | Reuse a seed to reproduce a take while changing the prompt. |
What it learns from a reference clip
A reference clip is richer than a still: it carries motion, camera movement, pacing and a visual style across time. Wan 2.6 uses those signals to generate a new shot that feels like it belongs in the same sequence. Give it a handheld documentary clip and the output inherits the handheld rhythm; give it a locked-off product turntable and the output inherits the clean, controlled camera.
This makes the endpoint useful for matching an existing campaign, extending a visual language across subjects, or testing a new scene in a style that already works. It is not a video editor and does not modify the reference; it creates a new clip with the same feel.
- Style transfer from the reference
- Motion and pacing transfer
- Camera language and framing
- 1 or 2 reference videos
- New scene, new subject, same feel
WAN 2.6 Reference-to-Video pricing
Billing is per output second by resolution. Reference clips do not add their duration to the bill, so a 10-second output costs the same whether the reference clip is short or long. There is no per-reference fee.
- Billed on output duration only
- Reference clip length is not added to the bill
- 5 or 10-second output
- Landscape and portrait sizes
| Resolution | Credits/s | 5s | 10s |
|---|---|---|---|
| 720p | 50 | 250 | 500 |
| 1080p | 80 | 400 | 800 |
Sizes, audio and controls
Supported sizes are 1280×720 and 1920×1080 in landscape, 720×1280 and 1080×1920 in portrait. Pick the size that matches the reference's framing; a portrait reference driving a landscape output loses much of its composition.
Audio is generated natively and included in the per-second rate. An optional audio upload of up to 15MB can guide the sound bed. Reference videos can be up to 100MB each, and the endpoint supports a seed for repeatable results.
- 1280×720 or 1920×1080 landscape
- 720×1280 or 1080×1920 portrait
- Native audio included
- Optional audio upload up to 15MB
- Up to 100MB per reference video
- Seed support for repeatability
Prompting against a reference
Describe the new subject and scene; let the reference carry the style. If you repeat the reference's content in the prompt, the model tends to copy rather than transfer. A useful structure is: new subject, new setting, one action, then a short note about which quality to inherit — pacing, camera, light or grade.
Two references are useful when you want to separate concerns: one for camera and motion, one for colour and texture. If the output feels like a copy of the reference, reduce the similarity of the prompt and remove the second clip.
- Describe the new scene, not the reference scene
- Name the quality to inherit
- Use two references for separate concerns
- Reduce prompt similarity if the output copies
- Keep the output size close to the reference framing
Limitations and alternatives
The endpoint takes video references only — no reference images — and caps at 10 seconds. It also has a narrower resolution range than Wan 3.0: 720p and 1080p, with no 480p or 4K. If you need image references or 30-second output, use Wan 3.0 Reference-to-Video.
For a single still, a standard image-to-video model is simpler. For prompt-driven edits to existing footage, use Wan 3.0 Video Edit or Kling O3 Pro Video Edit.
- Video references only
- 10-second maximum
- No 480p or 4K
- For image references, use Wan 3.0
- For edits, use a video-edit endpoint
Prompt examples for WAN 2.6 Reference-to-Video
- Documentary style transfer: A fisherman repairs a net on a wooden dock at dawn, camera handheld at eye level, natural light, muted coastal palette. Match the pacing and camera texture of the reference. Gulls, water, rope. Inherit the handheld rhythm from the reference.
- Product turntable: A glass perfume bottle rotates on a dark reflective surface, a narrow highlight sweeping across the label, locked camera. Match the reference's clean studio look. Soft room tone, glass chime. Use a locked-off reference for a controlled result.
- Two-reference split: A cyclist climbs a coastal road at sunset, camera tracks from a car, warm grade. Use Video 1 for camera movement and Video 2 for colour and texture. Wind, tires, distant surf. One reference for motion, one for look.
- Portrait framing: A potter shapes a bowl at a wheel, hands in frame, warm window light, shallow depth of field. Match the portrait framing and pace of the reference. Clay, wheel hum, quiet room. Keep the output size portrait.
- Portrait to landscape caution: A chef plates a dish in a busy kitchen, camera moves along the counter, overhead light, quick pace. Match the reference's energy. Sizzling, knives, kitchen ambience. A landscape output needs a landscape reference.
- Audio-guided: A drummer performs a slow fill in a concrete room, camera circles, hard sidelight, dust in the beam. Use the uploaded audio as the rhythm guide; drums, room reverb. Audio upload is optional and capped at 15MB.
Pro tips for WAN 2.6 Reference-to-Video
This endpoint transfers style and motion, not a specific identity. Cast it for look and movement.
- References teach style, not identity If you need a specific face or product, start from an image-to-video model instead. Reference-to-video is for look, pacing and camera language.
- Keep references short and consistent A few seconds of clear motion teaches more than a minute of mixed shots. Trim references to the behavior you want copied.
- 1080p is a 60 percent premium Eighty credits per second versus fifty. A 10-second 1080p clip costs 800 credits, so validate the scene at 720p first.
- Two references maximum A third clip rarely adds a new idea and often blurs the style. If two references conflict, decide which one wins before generating.
- Write the prompt without describing the references Do not repeat what the reference already shows. Describe the new scene and let the model carry the grammar across.
Frequently asked questions
How much does WAN 2.6 Reference-to-Video cost?
It costs 50 credits per second at 720p and 80 at 1080p. A 5-second 720p clip is 250 credits; a 10-second 1080p clip is 800.
How many reference videos can I use?
One or two. Reference images are not accepted on this endpoint.
Is the reference video length billed?
No. Billing is on the output duration only, so a 10-second output costs the same regardless of reference length.
What resolutions and sizes are supported?
1280×720 and 1920×1080 in landscape, 720×1280 and 1080×1920 in portrait.
How long can the output be?
5 or 10 seconds.
Does it generate audio?
Yes. Native audio is included, and an optional audio upload of up to 15MB can guide the sound bed.
What does the model learn from the reference?
Style, motion, framing, camera language and pacing. It creates a new shot rather than editing the reference.
Can I get repeatable results?
Yes. The endpoint supports a seed, so the same prompt, references and seed can be reproduced.
Compare with other reference-to-video models
- Wan 3.0 Reference-to-Video
- Kling Reference-to-Video
- WAN 2.6 Text-to-Video
- Wan 3.0 Video Edit
- Kling O3 Pro Video Edit
- Browse all AI video models
Specifications
- Provider: Wavespeed
- Model type: Video to video
- Credit pricing: 720p: 50 credits · 1080p: 80 credits
- Capabilities: new, cinematic, starter, reference-based, multi-shot
About Conut.ai
Conut.ai is an AI image, video and audio generation platform with 50+ creative tools. It supports text-to-image, image-to-video, face swap, motion control, relighting, inpainting and video effects through models including Nano Banana Pro, Seedream, GPT Image, Sora 2, Kling, WAN and Seedance. Start free — no credit card required.