Reference to VideoKeep the Character, Product and Set
Attach three or more images below and they become references. Tag them in the prompt as @Image 1, @Image 2 and say what each one is; the model keeps them and builds a new shot around them. Seedance 2.5 by default; 6 more models take references.
Models, limits and prices read from the live generator on .
Image to video starts the clip on your picture. Reference to video does not: it reads your images for identity, wardrobe, product details, set and style, and generates a new shot in which they appear, with the camera, poses and action you describe. That is how a character stays the same across several clips, a product lands in a scene it was never photographed in, or a whole cast, costume and location go into one shot.
Reference to video means you upload pictures of who and what should appear, such as a character, a product, an outfit or a place, and the AI builds a new shot around them instead of starting the clip on one picture. On SeedVideo AI, 7 models take references: Seedance 2.5 (up to 30 images and 10 video clips), Wan 3.0 (10 images), Seedance 2.0, 2.0 Fast, 2.0 Mini and MiniMax H3 (9 images each), and Kling O3 (one video clip). Reference images add no cost on Seedance: you pay the normal rate, from 1.16 credits per second (Seedance 2.0 Mini, 480p) to 6.10 on Seedance 2.5 at 720p. Free sign-up credits run Seedance 2.0 Mini.
An official Seedance 2.5 example: 18 references, one 30-second shot
ByteDance Seed published this case with the Seedance 2.5 announcement. Eighteen images went in as references, none of them a frame of the video: the concert hall, the pianist, a cello, a violin, the lead vocalist, five orchestra players, four choir members and four audience members. The instruction gives each image or group of images a role and then directs one continuous concert scene.
What carried over: the hall’s gilded dome and organ from the first image, the pianist’s face and tuxedo, the lead vocalist’s navy off-shoulder gown and hairstyle from the three-view sheet in @Image 5, and the audience members’ outfits. The singer’s reference shows her from the front, side and back, which is why she stays recognisable as the camera circles the stage.
- Model: Seedance 2.5 · Inputs: 18 reference images + instruction
- Duration: 30 seconds · Aspect ratio: 16:9 · Audio: generated with the picture
- The same job on SeedVideo AI (18 images, 30 s on Seedance 2.5): 92 credits at 480p, 183 credits at 720p, 453 credits at 1080p at the catalogue rate; the images add nothing
The instruction first sets format and look, then assigns roles (“Use @Image 1 for the venue”, “The lead vocalist must strictly follow @Image 5”, “Reference @Images 11 to 14 for the choir”), places everyone on the stage, and only then describes the camera and the action from the opening wide shot to the applause.
Full instruction (English, verbatim from the announcement)
A 30-second concert sequence in 16:9 landscape, with cinematic realism, authentic concert hall lighting and shadows, warm golden stage lighting, and the atmosphere of a formal classical concert. Use @Image 1 for the venue. Reference @Image 2 for the pianist. Reference @Image 3 for the cello. Reference @Image 4 for the violin. The lead vocalist must strictly follow @Image 5. Reference @Images 6 to 10 for the rest of the orchestra. Reference @Images 11 to 14 for the choir. Reference @Images 15 to 18 for the audience seating. The lead vocalist walks from center stage toward the front edge. The pianist is positioned by the piano. The orchestra is arranged on both sides and toward the rear. The choir stands at the back of the stage. Open with a high-angle wide shot of the full concert hall. The pianist strikes the keys, and the lead vocalist steps into the spotlight and begins singing. The camera naturally moves across the violin, cello, and orchestra as they perform together, with the violin feeling bright and the cello warm. In the latter part, the choir joins in. The lead vocalist briefly makes eye contact with front-row audience members, who respond with a smile and a slight nod. In the closing shot, the camera pulls back. The singing ends, and the audience joins in the applause.
Video, reference images and instruction: ByteDance Seed, from the Seedance 2.5 announcement. Shown here from the output onward, without the prompt cards. Source. Open this instruction in the generator and attach your own images in the same order; nothing starts until you press Generate.
How to make a video from reference images
- 1
Collect one clean image per role
One image for each character, one for the product or outfit, one for the place or style. Plain backgrounds and a clear view of the face or object work best; a front, side and back sheet holds a character through camera moves. Seedance 2.5 takes up to 30 images, Wan 3.0 up to 10, Seedance 2.0 models and MiniMax H3 up to 9.
- 2
Attach them as references
Attach three or more images and the generator treats them as references. With one or two images, click the image’s Start or End badge and choose “Use as reference image”. Adding a video clip on Seedance, Wan 3.0 or MiniMax H3 also switches the set to references.
- 3
Tag each image in the prompt
Type @ in the prompt to insert @Image 1, @Image 2 and so on, numbered in the order the images are attached; video clips are numbered after the images. Say what each one is for: “the woman from @Image 1”, “the jacket in @Image 3”.
- 4
Describe the new shot, then generate
Write the action, the camera and the light of the shot you want, and what must not change. Pick length and resolution; the composer shows the exact credit cost before you start, and failed or refused runs are not charged.
Starter instruction (three images: a person, a place, a product):
The woman from @Image 1 walks into the café from @Image 2, sits down and lifts the cup from @Image 3 toward the camera. Keep her face, hair and clothes exactly as in @Image 1 and the cup’s shape and label as in @Image 3. Handheld medium shot, warm morning light.
Models that support reference to video
Catalogue credits per second, before plan discounts. With images the rate is the model’s normal one; with a reference clip it is the rate the generator charges on the seconds shown.
| Model | References | Output | Credits per second | Who can run it |
|---|---|---|---|---|
| Seedance 2.5 | References:Up to 30 images and 10 video clips (2–30 s each) | Output:4–30 s | Credits per second:Images: 3.05 at 480p · 6.10 at 720p · 15.10 at 1080pVideo: 2.12 at 480p · 4.58 at 720p · 10.21 at 1080p, per second of clip + output | Every plan, every credit pack |
| Seedance 2.0 | References:Up to 9 images and 3 video clips (2–15 s each) | Output:4–15 s | Credits per second:Images: 2 at 480p · 4 at 720p · 10 at 1080pVideo: 1.41 at 480p · 3.05 at 720p · 6.88 at 1080p, per second of clip + output | Every plan, every credit pack |
| Seedance 2.0 Fast | References:Up to 9 images and 3 video clips (2–15 s each) | Output:4–15 s | Credits per second:Images: 1.74 at 480p · 3.75 at 720p · 8.47 at 1080pVideo: 1.06 at 480p · 2.29 at 720p · 5.16 at 1080p, per second of clip + output | Every plan, every credit pack |
| Seedance 2.0 Mini | References:Up to 9 images and 3 video clips (2–15 s each) | Output:4–15 s | Credits per second:Images: 1.16 at 480p · 2.50 at 720pVideo: 0.71 at 480p · 1.53 at 720p, per second of clip + output | Free credits, every plan, every credit pack |
| Wan 3.0 | References:Up to 10 images and 5 video clips (1–15 s each) | Output:2–30 s | Credits per second:Images: 1.20 at 480p · 2.40 at 720p · 4.80 at 1080pVideo: 1.20 at 480p · 2.40 at 720p · 4.80 at 1080p, per second of clip + output | Every plan, every credit pack |
| MiniMax H3 | References:Up to 9 images and 3 video clips (2–15 s each) | Output:5–15 s | Credits per second:Images: 4.55 at 2K (+1.4 credits per image past 5)Video: 4.55 at 2K, per second of clip + output | Pro and Premium plans, every credit pack |
| Kling O3 | References:One video clip (3–10 s); no reference images | Output:4–10 s | Credits per second:Video: 2.50 at 720p · 4 at 1080p, per output second; the clip is not billed | Every plan, Creator pack and up |
Resolution is also capped by plan: 720p on free credits and, for Seedance, on Lite; 1080p for Seedance on Pro, Premium and credit packs. Kling motion control also takes a character image and a clip, but it copies the clip’s movement rather than building a new shot, so it is not listed here.
What a reference to video run costs on each plan
Credits the generator charges on each plan today (plan discounts included, as a new subscriber), and the dollar value of those credits at the plan’s price per credit. Taxes not included. Failed or refused runs are not charged.
| Plan | Seedance 2.0 Mini, 3 images, 4 s, 480p | Seedance 2.5, 3 images, 10 s, 720p | Seedance 2.5, one 5 s reference clip, 5 s, 720p |
|---|---|---|---|
| Free sign-up credits$0 · 5 credits once | Seedance 2.0 Mini, 3 images, 4 s, 480pFree (5 of 5 credits) | Seedance 2.5, 3 images, 10 s, 720pNot on this plan | Seedance 2.5, one 5 s reference clip, 5 s, 720pNot on this plan |
| Lite, monthly$14.90/mo · 100 credits/mo | Seedance 2.0 Mini, 3 images, 4 s, 480p$0.75 (5 credits) | Seedance 2.5, 3 images, 10 s, 720p$8.20 (55 credits) | Seedance 2.5, one 5 s reference clip, 5 s, 720p$6.11 (41 credits) |
| Lite, billed yearly$13/mo · 100 credits/mo | Seedance 2.0 Mini, 3 images, 4 s, 480p$0.65 (5 credits) | Seedance 2.5, 3 images, 10 s, 720p$7.15 (55 credits) | Seedance 2.5, one 5 s reference clip, 5 s, 720p$5.33 (41 credits) |
| Pro, monthly$29/mo · 200 credits/mo | Seedance 2.0 Mini, 3 images, 4 s, 480p$0.58 (4 credits) | Seedance 2.5, 3 images, 10 s, 720p$6.24 (43 credits) | Seedance 2.5, one 5 s reference clip, 5 s, 720p$4.64 (32 credits) |
| Pro, billed yearly$14.50/mo · 200 credits/mo | Seedance 2.0 Mini, 3 images, 4 s, 480p$0.29 (4 credits) | Seedance 2.5, 3 images, 10 s, 720p$3.12 (43 credits) | Seedance 2.5, one 5 s reference clip, 5 s, 720p$2.32 (32 credits) |
| Premium, monthly$149/mo · 1,300 credits/mo | Seedance 2.0 Mini, 3 images, 4 s, 480p$0.34 (3 credits) | Seedance 2.5, 3 images, 10 s, 720p$4.93 (43 credits) | Seedance 2.5, one 5 s reference clip, 5 s, 720p$3.67 (32 credits) |
| Premium, billed yearly$99/mo · 1,300 credits/mo | Seedance 2.0 Mini, 3 images, 4 s, 480p$0.23 (3 credits) | Seedance 2.5, 3 images, 10 s, 720p$3.27 (43 credits) | Seedance 2.5, one 5 s reference clip, 5 s, 720p$2.44 (32 credits) |
| Starter credit pack$29.90 one-time · 200 credits | Seedance 2.0 Mini, 3 images, 4 s, 480p$0.75 (5 credits) | Seedance 2.5, 3 images, 10 s, 720p$9.12 (61 credits) | Seedance 2.5, one 5 s reference clip, 5 s, 720p$6.88 (46 credits) |
| Creator credit pack$49.90 one-time · 375 credits | Seedance 2.0 Mini, 3 images, 4 s, 480p$0.67 (5 credits) | Seedance 2.5, 3 images, 10 s, 720p$8.12 (61 credits) | Seedance 2.5, one 5 s reference clip, 5 s, 720p$6.12 (46 credits) |
“Not on this plan” means the plan cannot run that model or resolution: free credits cover Seedance 2.0 Mini only. On Seedance a reference clip is billed at a lower rate per second than images-only output, but its own seconds count too. See pricing for every plan and pack.
Reference to video specs
- Models
- Seedance 2.5, 2.0, 2.0 Fast and 2.0 Mini; Wan 3.0; MiniMax H3; Kling O3 (reference video only)
- Reference images
- Up to 30 on Seedance 2.5, 10 on Wan 3.0, 9 on Seedance 2.0, 2.0 Fast, 2.0 Mini and MiniMax H3. Seedance: JPEG, PNG or WebP up to 30 MB each
- Reference video
- MP4 or MOV. Seedance 2.5: up to 10 clips of 2–30 s, 200 MB each; Seedance 2.0 models and MiniMax H3: up to 3 clips of 2–15 s, 50 MB; Wan 3.0: up to 5 clips of 1–15 s, 100 MB; Kling O3: one clip of 3–10 s
- Output length
- 4–30 s on Seedance 2.5, 2–30 s on Wan 3.0 (reference clips plus output stay within 30 s), 4–15 s on Seedance 2.0 models, 5–15 s on MiniMax H3, 4–10 s on Kling O3
- Resolution
- 480p–1080p on Seedance 2.5, 2.0, 2.0 Fast and Wan 3.0 (Seedance 1080p needs Pro, Premium or a credit pack); 480p–720p on Seedance 2.0 Mini; 2K on MiniMax H3; 720p–1080p on Kling O3
- Price
- Images: the model’s normal rate, nothing extra (MiniMax H3 adds 1.4 credits per image past five). Video: Seedance bills clip + output seconds at a lower reference rate, Wan 3.0 and MiniMax H3 clip + output seconds at their normal rate, Kling O3 output seconds only
- Audio
- Seedance and Wan 3.0 generate sound with an on/off switch at the same price; MiniMax H3 returns sound with no switch; Kling O3 keeps the reference clip’s own sound
- Watermark
- None, on any plan or free credits
- Generation time
- Median about 3 minutes on Seedance 2.5 and under 2 minutes on Seedance 2.0 Mini for reference runs over the last 30 days
Tips for consistent characters and products
- Give every image one job and say it in the prompt. “The face and hair from @Image 1, the dress from @Image 2” beats uploading five photos and hoping the model sorts them out.
- For a consistent character across several clips, reuse the exact same reference images and the same tag wording in every prompt; change only the action and the camera.
- A front, side and back view of the same character (a three-view sheet in one image works) holds identity and costume better than a single portrait.
- Shoot products on a plain background and name what must survive: “keep the label, the cap and the bottle shape exactly as in @Image 3”.
- Don’t describe the face or outfit in words as well; text that disagrees with the picture pulls the result away from it.
- Real people’s faces, celebrities and copyrighted characters can be refused under the model providers’ content rules. A refused run is not charged.
When to use another tool
- The clip must start on your exact picture
- Image to video. The image becomes the first frame, so the composition is fixed, not just the identity.
- Both the opening and the closing picture are fixed
- First & last frame to video. Two images pin both ends of the shot on 11 models; the model fills in the motion.
- A character has to copy a dance or move from a clip
- Kling 2.6 Motion Control. One character image plus one motion video: the character performs the clip’s movement.
- You have the footage and want to change something in it
- Video to video. Replace a person, an object or the setting in an existing clip and keep its camera and timing.
- Your clip ends too early
- AI video extender. Continues an existing video from its last frame for 4 to 30 seconds on Seedance 2.5.
Reference to video FAQ
- What is reference to video (R2V)?
- Reference to video is a generation mode where your images, and on some models video clips, act as references for characters, products, wardrobe, setting and style. The AI generates a new shot in which those things appear, with the action and camera you describe. None of the images has to be a frame of the result.
- What is the difference between R2V and I2V?
- Image to video (I2V) animates your picture: the clip starts on it, so the composition, framing and pose are fixed and only motion is added. Reference to video (R2V) borrows what is in the pictures, such as a face, an outfit or a product, and puts it into a new composition. Use I2V when the first frame must match exactly, R2V when the character or product must match but the shot is new.
- How do I keep a character consistent across AI videos?
- Use the same reference images for every clip, ideally a front, side and back view plus a close-up of the face, and refer to them the same way in each prompt (“the man from @Image 1”). Change only the action, camera and setting between clips. Seedance 2.5 holds up to 30 references, so several characters, their costumes and the location can all go into one run.
- Which AI video models support reference images?
- On SeedVideo AI: Seedance 2.5 (up to 30 images), Wan 3.0 (10), and Seedance 2.0, 2.0 Fast, 2.0 Mini and MiniMax H3 (9 each). Kling O3 takes a reference video but no reference images. Seedance 1.5 Pro, Kling 3.0, Veo 3.1 and Sora 2 take a first frame, not a reference set.
- Do reference images cost extra?
- Not on Seedance or Wan 3.0: a run with reference images costs the same as a text-to-video run of the same length and resolution. MiniMax H3 adds 1.4 credits for each image past the fifth. A reference video is different: most models also bill its seconds, as the model table shows.
- Can I try reference to video for free?
- Yes. New accounts get free sign-up credits with no card, which cover a 4-second 480p reference run on Seedance 2.0 Mini with up to 9 images. Seedance 2.5, Wan 3.0, MiniMax H3 and Kling O3 need a plan or a credit pack. There is no watermark on any video.
- Can I use a video as a reference?
- Yes, on six models. Seedance 2.5 takes up to 10 clips of 2–30 s, Seedance 2.0 models and MiniMax H3 up to 3 clips of 2–15 s, Wan 3.0 up to 5 clips of 1–15 s, and Kling O3 one clip of 3–10 s. A clip can guide motion, camera or the look of a scene, and it can be combined with reference images on every model except Kling O3.
- Why does my character still change between shots?
- Usually the prompt describes the person in words that conflict with the image, the reference is small or partly hidden, or several references compete for the same role. Drop the written description of the face, use a clear front view, give each image one job, and keep the same images for every shot.
SeedVideo AI is an independent platform and is not affiliated with ByteDance, Kuaishou, Alibaba, MiniMax or the other model providers named here. The example above was published by ByteDance Seed and is credited to it.