Reference to video AI generates a clip from two or more images of the same person, product or place, so the subject looks identical from the first frame to the last while the prompt controls only the action. On Aura AI seven model families do it in the reference to video tool: Kling O3 Reference (720p, or 4K on Premium), Veo 3.1 Reference, Seedance 2.0 Reference with its Fast and Mini variants, Wan 2.7 Reference, Vidu Q2 Reference, Grok Reference and Gemini Omni Flash Reference. Your plan sets how many images you can attach per generation, from 2 on Starter up to 9 on Premium with Seedance.
In This Article
How reference to video differs from image to video
Image to video animates one picture: the upload becomes the first frame and the model moves it. Reference to video treats your images as identity anchors, not frames. The model reads the face, the outfit, the product's shape and the room from several angles, and then stages a new scene around that subject from the prompt. That is why a reference clip can open on a wide shot the photos never showed, and why a character can be reused across a series without drifting.
Two consequences follow. First, more angles beat more copies: a front view and a three-quarter view teach the model more than two near-identical selfies. Second, the prompt should not redescribe the subject. Describe the action, the camera and the sound, and let the references handle the look. Kling's own guidance for O3 recommends a front-facing and a side-profile photo to lock a face, and that holds across the models below.
Reference to video models compared
Limits are per generation, by plan (Starter / Pro / Premium). Credits are for an 8 second clip at the default resolution.
| Model | Reference images | Duration | Resolution | Audio | Credits (8 s) | Plan |
|---|---|---|---|---|---|---|
| Kling O3 Reference | 2 / 3 / 4 | 3 to 15 s | 720p | Yes | 8 | Every plan |
| Kling O3 4K Reference | 4 (Premium) | 3 to 10 s | 4K | Yes | 15 (10 s: 17) | Premium |
| Veo 3.1 Reference | 2 / 3 / 5 | 8 s | 720p, 1080p, 4K | Yes | 6 at 720p, 7 at 1080p, 11 at 4K | Every plan |
| Seedance 2.0 Reference | 2 / 5 / 9 | 4 to 15 s | 480p to 4K | Yes, stereo | 8 at 720p, 15 at 1080p, 30 at 4K | Starter+ |
| Seedance 2.0 Fast Reference | 2 / 5 / 9 | 4 to 15 s | 480p, 720p | Yes | 5 at 720p | Starter+ |
| Seedance 2.0 Mini Reference | 2 / 5 / 9 | 4 to 15 s | 480p, 720p | Yes | 4 at 720p | Starter+ |
| Wan 2.7 Reference | 2 / 3 / 5 | 2 to 10 s | 720p, 1080p | No | 6 at 720p, 7 at 1080p | Starter+ |
| Vidu Q2 Reference | 2 / 3 / 5 | 3 to 8 s | 540p, 720p, 1080p | No | 6 at 720p, 8 at 1080p | Every plan |
| Grok Reference | 2 / 3 / 7 | 3, 5, 8 or 10 s | 480p, 720p | Yes | 6 at 720p | Every plan |
| Gemini Omni Flash Reference | 2 / 3 / 5 | 3 to 10 s | 720p | Yes | 5 | Every plan |
Fractions round up to the next whole credit and only successful generations are charged. Kling and Vidu offer 16:9, 9:16 and 1:1, Gemini Omni Flash 16:9 and 9:16, Seedance six ratios, Wan 2.7 five and Grok seven.
What each model is good at
Kling O3 Reference and O3 4K Reference
Kling O3 is the omni line built for prompt adherence and multi-shot consistency, and its reference mode is the one to pick when the clip has to move: a character walking through a market, a product turning in hand. It runs 3 to 15 seconds at 720p with native audio for 6 credits plus 0.5 per second, with up to 4 images on Premium. The 4K version renders the same subject at 3840x2160 in 3 to 10 second clips for 12 credits plus 0.9 per second, Premium only.
Veo 3.1 Reference
Google's ingredient-style references on Veo 3.1 produce the most cinematic result of the group, with dialogue, effects and ambience generated alongside. Clips are a fixed 8 seconds at 720p, 1080p or 4K, from 2 images on Starter to 5 on Premium, at 5 credits plus 0.6 per second (7 credits at 1080p, 11 at 4K). It is the model behind the example in our original reference to video guide.
Seedance 2.0 Reference, Fast and Mini
Seedance 2.0 accepts the most references on Aura AI, 9 images on Premium and 5 on Pro, and upstream it mixes images with video clips and audio tracks, up to 12 files, with an @tag system to name each character. That makes it the choice for scenes with two or more consistent people, for choreography synced to a track, and for 4K reference clips from the Starter plan. Fast and Mini keep the same 9-image limit at 480p or 720p for 4 and 3 credits base, the cheapest reference generations in the catalog.
Wan 2.7 Reference
Wan 2.7 is the flexible all-rounder: 2 to 10 second clips at 720p or 1080p in five aspect ratios, 2 to 5 images by plan, 5 credits plus 0.6 per second. Reference mode is silent, so add sound in the editor, but it is quick and holds product shapes well.
Vidu Q2 Reference
Shengshu's Vidu Q2 Reference handles subjects, environments, costumes and props as separate references and composes them into a 3 to 8 second clip at 540p, 720p or 1080p. It is silent and short, but it is on every plan from 3 credits at 540p, a good place to test a character before spending Kling or Veo credits.
Grok Reference
xAI's Grok Reference allows the most images after Seedance, 7 on Premium, in 3, 5, 8 or 10 second clips at 480p or 720p with native audio, for 4 credits plus 0.5 per second. Seven aspect ratios cover formats the others skip.
Gemini Omni Flash Reference
Google's Gemini Omni Flash prices reference to video exactly like text to video: 3 credits up to 4 seconds and 0.3 per extra second, any length from 3 to 10 seconds at 720p with audio, 2 to 5 images by plan. A character-locked shot carries no premium, so it suits a series of short scenes with the same presenter.
How many reference images you need
- One person: two images, front and three-quarter or profile, same lighting, no filters. Starter allows two on every model, enough for a face.
- A product: three angles, front, side and detail, on a plain background. Pro (3 images) covers this on Kling, Veo, Wan, Vidu and Gemini; Seedance and Grok allow more.
- A person in a specific place or outfit: four to five images: two of the face, one full-body, one or two of the environment. Premium on Veo, Wan, Vidu, Gemini (5) or Kling O3 (4).
- Two or more characters together: Seedance 2.0 Reference on Pro (5) or Premium (9), naming each character with an @tag in the prompt, or Grok Reference on Premium (7).
Reference quality checklist
Sharp, well lit, no heavy filters, consistent outfit across the set, and no conflicting angles of different people. If a photo is not right, fix it in the AI image editor (remove a background, change a color, clean a logo) before uploading. Then describe only the action, camera and sound.
How to do it in Aura AI
- Open Reference to Video in the studio and sign in. Upload 2 to 9 images depending on your plan and the model; the tool shows the limit as you add them.
- Pick the model. Kling O3 for motion, Veo 3.1 for cinematic realism with dialogue, Seedance 2.0 for several characters or 4K, Gemini Omni Flash or Seedance Mini to test cheaply. The credit cost updates live.
- Describe the scene, not the subject. Example: the referenced woman walks into a sunlit cafe, orders a coffee and sits by the window, cafe chatter, she says "Finally", 16:9, 8 seconds.
- Generate, preview and download the MP4 without watermark. Reuse the same reference set for the next scene.
Keep your character consistent across every clip
Upload two to nine reference images and generate with Kling O3, Veo 3.1, Seedance 2.0, Grok, Vidu, Wan 2.7 or Gemini Omni Flash. HD, 4K, native audio, no watermark.
Open Reference to VideoFrequently Asked Questions
What is reference to video AI and how is it different from image to video?
Reference to video generates a new clip from two or more images that define a subject's identity, while the prompt controls the action, camera and sound. Image to video animates a single picture as the first frame. Because the references act as anchors rather than frames, the same character or product stays consistent across scenes and clips.
How many reference images can I upload?
It depends on your plan and the model. Starter allows 2 images on every model. Pro allows 3 on Kling O3, Veo 3.1, Wan 2.7, Vidu Q2, Grok and Gemini Omni Flash, and 5 on Seedance 2.0. Premium allows 4 on Kling O3, 5 on Veo 3.1, Wan 2.7, Vidu Q2 and Gemini Omni Flash, 7 on Grok and 9 on Seedance 2.0.
Which reference to video model keeps characters most consistent?
For a single person in motion, Kling O3 Reference is the usual pick, with Veo 3.1 Reference for the most cinematic look and dialogue. For two or more characters in one scene, Seedance 2.0 Reference with its @tag system and up to 9 images is the strongest option. A front view plus a profile in consistent lighting beats several near-identical photos.
Does reference to video generate audio?
Kling O3, Veo 3.1, Seedance 2.0 (including Fast and Mini), Grok and Gemini Omni Flash generate native audio in reference mode, so a referenced character can speak a quoted line with lip-sync. Wan 2.7 Reference and Vidu Q2 Reference produce silent clips; add sound to those in the AI video editor.
How much does reference to video cost on Aura AI?
An 8 second reference clip is 4 credits on Seedance 2.0 Mini, 5 on Seedance 2.0 Fast or Gemini Omni Flash, 6 on Grok, Vidu Q2 or Wan 2.7 at 720p, 7 on Veo 3.1 at 1080p, 8 on Kling O3 or Seedance 2.0 at 720p, and 15 on Kling O3 4K Reference.