Make two AI clips of the same character and the second one often comes out with a different face, a different outfit, or a different style. Reference-to-video fixes this: rather than describing your subject and hoping the model gets it right, you show it. Below: what it is, a workflow, prompts you can copy, and what it costs on Veemo.
What is reference-to-video?
With reference to video, you upload one or more images of a subject (a character, a product, an outfit, a place, or a visual style) and write a prompt for a brand-new scene. The model keeps the subject from your references and builds the motion, background, and camera work around it.
Reference-to-video vs image-to-video:
- Image-to-video animates one specific photo. That photo becomes the first frame, so you stay in that scene.
- Reference-to-video uses your images as a guide, not a first frame. The subject can appear in a whole new setting, pose, or camera angle, which is what you want for series, ad campaigns, and storyboards.
A quick rule: if the job is "make this picture move," use image-to-video. If it's "put this character or product somewhere new," use reference-to-video.
Which models support it on Veemo
When we checked the generator (October 2026), these models had a reference-to-video mode. Each one accepts different inputs:
| Model | Reference inputs | Length | Resolution |
|---|---|---|---|
| MiniMax H3 | Up to 9 images, plus optional videos and audio | 4-15 s | 768P or 2K |
| Seedance 2.5 | Images, videos and audio (up to 30 images) | 4-30 s | 480p or 720p |
| Veo 3.1 | 1-3 reference images | Set by the model | 720p or 1080p |
| Wan 2.6 | Up to 3 references | 5 or 10 s | 720p (1080p at 5 s) |
| Seedance 2.0 / Fast / Mini | A reference video plus up to 9 images | 4-15 s | Up to 720p (2.0 also goes higher) |
| Kling O3 | A source video plus up to 4 images | 3-10 s | 720p or 1080p |
If you only have still images, start with MiniMax H3, Seedance 2.5, or Veo 3.1. The Seedance 2.0 models and Kling O3 expect a video input as well.
Step-by-step workflow
1. Make or pick clean reference images. Use sharp, well-lit images where the subject fills a good part of the frame. If you don't have any yet, generate them with Nano Banana Pro. It supports multiple reference images and character consistency for up to 5 people, and outputs at 1K, 2K, or 4K. A front view, side view, and full-body shot on a plain background work well.
2. Choose a model. Pick based on your inputs (images only, or video too), clip length, and budget.
3. Upload your references. Add more images only when each one adds something, such as a side profile or the outfit.
4. Write the prompt. Say what each reference is, then describe the scene, action, camera, and light. The images already carry the identity, so keep the subject description short.
5. Set the format. Choose aspect ratio (9:16 for Reels and TikTok, 16:9 for YouTube), length, and resolution. The cost is shown before you generate.
6. Test short, then scale. Run a short, low-resolution test. If the identity holds, reuse the references and change only the scene prompt.
Copyable prompts
Change the bracketed parts to match your project.
1. Character consistency: new location
Use the woman from reference image 1 as the main character. Keep her face, hair, and red coat exactly the same. She walks through a busy night market in Tokyo, glancing at food stalls, warm lantern light, handheld camera following from behind, then she turns toward the camera and smiles.
2. Character consistency: series episode
Same character as the references: [short description]. Morning scene in a small sunlit kitchen. He pours coffee, looks out the window, and checks his phone. Medium shot, soft natural light, calm cozy mood, realistic style.
3. Product ad: hero shot
The product is the bottle from the reference images. Keep the label, cap, and color exactly as shown. The bottle stands on wet black stone, water droplets slide down the glass, slow 180-degree orbit around it, dramatic rim lighting, premium commercial look.
4. Product ad: lifestyle
A young woman holds the sneaker from reference image 1 and laces it up on a city rooftop at sunset, then jogs toward the camera. Keep the shoe's colors and logo unchanged. Vertical 9:16, energetic social ad style.
5. Style transfer
Match the visual style of reference image 2: [hand-painted watercolor / 1980s anime / claymation]. The character from reference image 1 rides a bicycle along a coastal road, wind in her hair, wide tracking shot. Keep the character's design consistent with reference image 1.
6. Multi-character scene
Reference image 1 is the man with the gray beard, reference image 2 is the girl with the yellow raincoat. They sit together on a park bench in light rain and share an umbrella. He tells a story, she laughs. Keep both characters looking exactly like their references. Medium two-shot, gentle camera drift.
7. Character plus product
The character from reference image 1 unboxes the headphones from reference image 2 at a desk, puts them on, and nods to the music. Keep both the character and the product design consistent. Bright home studio, YouTube review style.
Tips and common mistakes
- Messy references. Busy backgrounds, small faces, heavy filters, and sunglasses all weaken the result.
- Fighting the references in the prompt. If the reference shows a red coat and the prompt says "blue jacket," expect drift. Describe the scene, not a new look.
- Not labeling your images. With more than one reference, say "reference image 1 is..." so the model knows which subject is which.
- Extreme angles and lighting. Looking straight up, full profiles, and deep shadows are where details slip most.
- Long clips before tests. Check identity on a short clip before paying for 15 or 30 seconds.
- Changing everything at once. For a series, keep the references and style wording fixed and change only the action and location.
How much does it cost?
Veemo uses credits, and the generator shows the exact cost before you generate. Reference-to-video prices when we checked:
- MiniMax H3: 4 credits per second at 768P, 7 per second at 2K. A 4-second 768P clip costs 16 credits.
- Seedance 2.5: 22 credits per second at 480p, 47 per second at 720p (less when you add a reference video).
- Veo 3.1: 40 credits per generation at 720p, 50 at 1080p.
- Wan 2.6: 80 credits for 5 s at 720p, 150 for 5 s at 1080p.
Free credits: new verified accounts get 16 credits, valid for 30 days, with no payment details required. That covers a 4-second MiniMax H3 clip at 768P, which is plenty for a first test. Downloading a result requires a credit pack or plan.
Paid options: subscriptions start at $29.90/month (598 credits) and annual billing is 50% off. One-time credit packs start at $9.90 for 152 credits, and pack credits never expire.
FAQ
How many reference images can I upload?
It depends on the model: up to 3 on Veo 3.1 and Wan 2.6, up to 9 on MiniMax H3, and more on Seedance 2.5. The upload box shows the limit.
Can I use the same character in many different videos?
Yes, that's what it's for. Keep the same references and change the scene prompt for each clip.
Do I need a video as a reference?
No. MiniMax H3, Seedance 2.5, and Veo 3.1 work with images only. The Seedance 2.0 models and Kling O3 need a video input in this mode.
Can I use the videos commercially?
Veemo's reference-to-video page says the content you generate comes with full commercial rights and needs no attribution.
Ready to try it? Open Veemo reference-to-video, upload your best reference, and run a short test.
Top comments (0)