Kling Reference-to-Video
Kling Omni O1 Reference-to-Video keeps a character, prop or scene consistent across new shots. 3–10s, three ratios, 60 credits/s with images, 80 with video.
Kling Omni O1 Reference-to-Video builds a new shot around subjects you already have. Upload up to 7 reference images — or images plus a short reference video — and describe a new scene; the model keeps the character, prop or environment recognisable while changing the context. Clips run 3 to 10 seconds at 16:9, 9:16 or 1:1. Image references cost 60 credits per second; attaching a reference video raises the rate to 80.
- References: 7 images + video
- Duration: 3s – 10s
- Price: 60–80 credits/s
- Ratios: 16:9 / 9:16 / 1:1
How to use Kling Reference-to-Video
Upload the subject from several angles, describe the new scene, and keep the identity cues consistent across references.
- Collect clean reference images Use up to seven JPG or PNG files with one subject per role: face, outfit, prop or environment. Keep each image sharp and at least 300 px per side.
- Add a reference video only if motion matters A short clip can guide movement or camera style, but it raises the rate from 60 to 80 credits per second for the entire generation.
- Describe the new scene State what changes: location, action, lighting. Then restate what must stay: the same face, outfit and distinguishing marks.
- Set duration and ratio Choose 3–10 seconds and one of the three ratios. Shorter clips drift less when the reference set is complex.
- Compare identity across the clip Check the face and props at the start, middle and end. If identity slips, reduce the number of references and give each one a clearer role.
Settings reference
Reference-to-Video builds a new scene around supplied subjects. Image-only references bill at 60 credits/s; adding a reference video raises the whole run to 80.
| Setting | Value | Notes |
|---|---|---|
| Reference images | Up to 7 | JPG or PNG, at least 300 px per side, up to 10 MB each, aspect ratio 1:2.5–2.5:1. |
| Reference video | Optional, 1 clip | Up to 10s, 700–2160 px wide, 200 MB maximum. |
| Duration | 3–10s | Default 5s. Billed per output second. |
| Price | 60 / 80 credits/s | 60 with images only, 80 when a reference video is attached. |
| Aspect ratio | 16:9 / 9:16 / 1:1 | Default 16:9. |
| Audio | None | Silent output; add sound in post. |
What identity lock means in practice
The model is built to separate a subject from its original scene. Give it pictures of a person, a product or a location, describe a new situation, and it generates that subject in the new context while preserving the features that make it recognisable. This is different from image-to-video, which animates one frame, and from text-to-video, which invents the subject from the prompt.
The practical limit is that references establish appearance, not behaviour. Identity, clothing details and product shape hold well; a specific gesture or a precise camera move is better specified in the prompt or supplied with a reference video. Multi-angle images help the model understand the subject in three dimensions.
- Character consistency across new scenes
- Product shape and label retention
- Multi-angle reference support
- New environments around an existing subject
- Text prompts direct the action
Kling Reference-to-Video pricing
Billing is per output second and depends on the reference type. Image-only references cost 60 credits per second; a reference video raises the rate to 80 credits per second for the whole generation. There is no separate per-image fee, and resolutions follow the reference set rather than a selectable resolution menu.
- Up to 7 reference images
- One optional reference video
- Image-only references stay at 60 credits/s
- A video reference prices the entire clip at 80 credits/s
- No per-image surcharge
| Reference type | Credits/s | 3s | 5s | 10s |
|---|---|---|---|---|
| Images only | 60 | 180 | 300 | 600 |
| With reference video | 80 | 240 | 400 | 800 |
Reference requirements
Images must be JPG or PNG, at least 300 pixels on each side, no larger than 10MB each, and within a 1:2.5 to 2.5:1 aspect ratio. That rejects extreme panoramas and tall crops, so use a crop close to a normal photograph. Different angles of the same subject are more useful than several near-identical shots.
A reference video must be 3 to 10 seconds long, with a width between 700 and 2160 pixels and a file size up to 200MB. It is optional, but it is the only way to transfer motion or camera language rather than appearance. Without any references the endpoint accepts only 5 or 10-second clips; with references, the full 3–10 second range is available.
- JPG or PNG, ≥300px per side, ≤10MB per image
- Aspect ratio between 1:2.5 and 2.5:1
- Up to 7 images
- Video: 3–10s, width 700–2160px, ≤200MB
- Without references: 5s or 10s only
- With references: any whole second from 3 to 10
Writing a reference prompt
Describe the new scene, not the reference. The model already knows what the subject looks like, so spend the prompt on where it is, what it is doing and how the camera behaves. A useful structure is: subject action, environment, camera, light. Keep the subject description short and consistent with the images.
If the output drifts, remove references before adding more. Conflicting angles, different lighting or two versions of a costume force the model to average them. Start with the minimum set that identifies the subject, confirm the result, then add detail only where it is missing.
- Describe the new scene, not the reference subject
- Subject action, environment, camera, light
- Use the fewest references that identify the subject
- Remove conflicting references first
- Use a video reference for motion, not identity
Limitations and alternatives
Kling Reference-to-Video caps at 10 seconds, generates no audio and supports only three aspect ratios. It is also not a video-edit endpoint: it creates a new performance rather than modifying existing footage. For edits, use Kling O3 Pro Video Edit; for longer clips, use Wan 3.0 Reference-to-Video, which reaches 30 seconds and adds audio.
If you only need to animate a single still, Kling Image-to-Video or Seedance I2V is simpler and cheaper. Reference-to-video is justified when identity must survive a change of scene.
- 10-second maximum
- No generated audio
- Three aspect ratios
- Not a video editor
- For 30 seconds, use Wan 3.0 Reference-to-Video
Prompt examples for Kling Reference-to-Video
- Character in a new city: The woman from the references walks through a neon-lit Tokyo alley at night, camera tracks beside her, rain on the pavement, cool blue and magenta light. Rain, footsteps, distant conversation. Use 2–3 angles of the same face.
- Product on location: The bottle from the references sits on a weathered wooden table at a beach cafe, morning light, waves blurred in the background, slow push-in. Gulls, surf, quiet cafe ambience. Product references work best on plain backgrounds.
- Costume change: The character from the references now wears a formal black suit and steps out of a car at a gallery opening, camera tracks forward, flashbulbs, cool evening light. City ambience, camera shutters, footsteps. Describe new wardrobe explicitly.
- Motion reference: Use the reference video for the camera path. The character from the images walks through a concrete underpass, long shadows, hard afternoon light. Footsteps, wind, distant traffic. A video reference raises the rate to 80/s.
- Environment consistency: The same apartment from the references, now at night, lamps on, city lights through the window, camera pans slowly across the room. Quiet room tone, distant traffic. Environmental references hold layout and decor.
- Group shot: Three characters from the references stand together on a rooftop at dusk, wind moving their coats, camera orbits slowly, cold blue sky. Wind, distant city, fabric movement. Seven images is enough for a small cast.
Pro tips for Kling Reference-to-Video
Reference-to-video is for changing the context without losing the subject. More references are not automatically better.
- Use three angles, not seven duplicates Front, three-quarter and profile teach the model more about a face than seven near-identical photos. Use the remaining slots for outfit or props.
- Separate subject from scene Reference the person with images and describe the new location in text. Mixing location photos into a character set dilutes identity.
- Avoid reference video unless needed It raises the per-second price by a third. Attach it when camera movement or performance style is the point, not for a static scene.
- Name the invariants Say explicitly that the face, hair and outfit must not change. Without that constraint, the model may redesign the subject for the new scene.
- Keep clips short with large reference sets Three to five seconds is enough to establish a character in a new place. Longer clips give identity more chances to drift.
Frequently asked questions
How much does Kling Reference-to-Video cost?
It costs 60 credits per second with image references and 80 credits per second when a reference video is attached. A 5-second image-only clip is 300 credits.
How many reference images can I upload?
Up to 7 JPG or PNG images, at least 300 pixels per side and no more than 10MB each.
Can I upload a reference video?
Yes. The video must be 3–10 seconds, 700–2160 pixels wide and at most 200MB. Attaching one raises the rate to 80 credits per second.
How long can the output be?
With references, any whole second from 3 to 10. Without references, only 5 or 10 seconds.
Does it generate audio?
No. The endpoint produces silent video; add sound in post or use a model with native audio.
What aspect ratios are supported?
16:9, 9:16 and 1:1.
Can it edit an existing video?
No. It creates a new shot around your references. Use Kling O3 Pro Video Edit to modify existing footage.
How does it compare with Wan 3.0 Reference-to-Video?
Kling caps at 10 seconds and is silent; Wan 3.0 reaches 30 seconds, accepts image, video and audio references, and generates audio at 23/45/90 credits per second.
Compare with other reference-to-video models
- Kling O3 Pro Video Edit
- Wan 3.0 Reference-to-Video
- WAN 2.6 Reference-to-Video
- Kling Image-to-Video
- Kling O3 Pro Text-to-Video
- Browse all AI video models
Specifications
- Provider: Wavespeed
- Model type: Video to video
- Credit pricing: 60 credits per second
- Capabilities: new, popular, 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.
Explore more AI tools
- WAN 2.6 Reference-to-Video
- HappyHorse 1.0 Text-to-Video
- HappyHorse 1.0 Image-to-Video
- HappyHorse 1.0 Video Edit
- HappyHorse 1.0 Video Extend
- Seedance 2.0 Text-to-Video
- Seedance 2.0 Image-to-Video
- Seedance 2.0 Fast Text-to-Video
- Seedance 2.0 Fast Image-to-Video
- Smart Crowd Remove
- Image Watermark Remover
- Image Background Remover