Reference-to-Video Prompt Workflows That Hold a Story Together

Reference-to-video means giving the model source images to build around, so a character, product or room stays recognisable from clip to clip. On VdoBloom the dedicated model for this is HappyHorse 1.1 Reference-to-Video, which builds new scenes around your references across 4 to 15 second clips in nine aspect ratios; Seedance 2.5 also accepts reference images, videos and audio and charges a lower per-second rate when a reference video is attached. The workflow that holds a story together is the same either way: lock the reference, describe the scene rather than the character, and change one thing per clip.

Updated October 2026.

What reference-to-video actually does

Text-to-video and image-to-video both build a clip from a fixed starting point — a prompt, or a single frame. Reference-to-video is different: you supply one or more images of something that must stay the same — a face, a jacket, a product, a living room — and the model builds a new scene that includes it. The clip is not an animation of your image; it is a new shot that happens to contain the thing in your image.

That distinction is what makes it useful for stories. With image-to-video you can only ever move the frame you uploaded, so every shot has to exist as a still first. With reference-to-video you can hold the character constant while the scene changes completely — the same person on a bus, then in a kitchen, then on a rooftop — which is what a story actually needs.

The two reference workflows on VdoBloom

HappyHorse 1.1 Reference-to-Video

This is the purpose-built option. It builds new scenes around your reference images and keeps subjects consistent across 4 to 15 second clips, at 720p or 1080p, in nine aspect ratios: 16:9, 9:16, 1:1, 4:3, 3:4, 4:5, 5:4, 9:21 and 21:9. That ratio list is unusually wide, and it covers both the ultrawide 21:9 and the tall 9:21, which means the same character can carry a cinematic shot and a full-length vertical one from the same reference. Pricing is about 9 credits per second at 720p and roughly 11.75 per second at 1080p, so a 5-second 720p clip is 45 credits.

Seedance 2.5 with references attached

Seedance 2.5 runs 4 to 30 second clips at 480p, 720p or 1080p in six aspect ratios, and it accepts reference images, reference videos and reference audio. The pricing detail worth knowing is that attaching a reference video makes it cheaper per second — roughly 6.8 credits per second at 480p, 15.2 at 720p and 38 at 1080p with a reference video, against about 11.2, 25.2 and 63.2 without one. If your workflow already includes a reference clip, that is a substantial saving on the same generation.

The other model worth knowing about is Kling 3.0, which runs 3 to 15 seconds across standard, pro and 4K tiers from about 5.6 to 26.8 credits per second, with audio adding to the rate. It is not a reference model in the same sense, but its motion control is what you reach for when the requirement is that a specific movement is reproduced rather than that a specific face is preserved.

Six prompts built for reference-to-video

Notice that none of them describe the character. That is deliberate — the reference image is doing that job, and re-describing the subject in words gives the model a competing description to blend with the picture.

Scene change, character constant. “The character steps through a glass door onto a wet city street at night, neon reflecting in the pavement, the camera follows from behind at shoulder height, the character’s outfit unchanged, a slow walk of about three paces.”
Product in motion. “The bottle rests on a polished stone surface while the camera arcs slowly around it, a single soft key light travelling across the label, the bottle itself completely still, reflective floor, no other objects.”
Same room, different light. “A wide shot of the same living room at dawn, curtains open, cool blue light, everything in place exactly as before, camera locked off, nothing moving but the curtains.”
Character at a table. “The character sits at a cafe table by a window, lifts a cup with both hands and takes a sip, camera at eye level slightly to the left, warm morning light across the table, no other people in frame.”
Reference video as a guide. “Reproduce the walking pace and body movement of the reference clip exactly, in a new setting: a covered market aisle, midday light through the roof slats, camera tracking at the same speed, the character’s clothing unchanged.”
Two characters, two references. “Two people stand facing each other in a doorway, one offering a folded document, hands visible and still at the moment of the handover, camera at chest height between them, flat overcast daylight, the background unchanged.”
Vertical version of the same shot. “The character walks toward the camera down a narrow alley, full length, tall vertical framing with sky above the buildings, late afternoon light, camera retreating at walking pace, nothing else in frame.”

The workflow that holds a story together

Prompts are the easy part. This is the sequence that stops a set of clips looking like a set of clips.

Build the reference first, properly. A clean, evenly lit, front-facing image. If the reference is soft or shadowed, every clip built from it inherits the problem.
Write a one-line bible. Character, wardrobe, and setting, in fixed wording. Reuse those exact words in every prompt, including the ones where they feel redundant.
Describe the scene, not the subject. The reference supplies the subject. Your prompt supplies the location, the action, the camera and the light.
Change one variable per clip. Keeping the character and the wardrobe fixed, change only the location between clip one and clip two. Then change only the camera angle for clip three. Two changes at once and you cannot tell which one broke the consistency.
Match the aspect ratio to the delivery. Nine ratios on HappyHorse 1.1 means you never need to crop a horizontal performance into a vertical one. Decide where the clip is going before you generate, not after.
Draft at the cheap setting. Generate at 720p to check that the character held and the action reads, then re-run the ones that work at 1080p.
Export the last frame when you need a hand-off. For a hard continuity point, take the final frame of a clip and use it as the first frame of the next generation. It is a stronger guarantee than the reference alone.

Where this fits alongside the other approaches is worth being clear about. Text-to-video is for shots with no fixed subject. Image-to-video is for making one photograph move. Reference-to-video is for keeping a subject recognisable across shots that are otherwise unrelated — and it is the only one of the three that scales to a story. AI video character consistency covers the underlying problem in more depth, and how to keep a character consistent in AI images handles the still side, which you should get right before you animate anything.

Common mistakes

Re-describing the subject in the prompt. The model then has two descriptions to reconcile — your reference image and your words — and the result is neither. Describe the scene only.
Using a busy reference. A reference showing three people in a crowded room makes the model guess which one to preserve. Crop to one subject per reference.
Changing wardrobe without saying so. If a character changes clothes between clips, that is a deliberate story beat and you should state it. If you do not, the model will change it for you.
Asking for a cut. One generation is one continuous shot. Multi-shot sequences are built in an editor, not in a prompt.
Ignoring the reference-video discount on Seedance 2.5. If you already have a clip that shows the movement you want, attaching it lowers the per-second rate instead of raising it.
Assuming the same reference works on every model. Reference-image limits and behaviour differ by model. Check what the model accepts before you plan a scene around it.

What it costs

Reference-to-video is priced per second, and the cost is shown before every run. HappyHorse 1.1 Reference-to-Video is about 9 credits per second at 720p and roughly 11.75 at 1080p, so a 5-second 720p shot is 45 credits. Seedance 2.5 is cheaper with a reference video attached than without. Plans start at $15/month for 300 credits, or $10/month billed yearly at $0.05 per credit, and one-time packs start at $10.39 for 200 credits that never expire — which suits a story shot over several sittings. India plans start at ₹199/month. New accounts get 10 free credits with no card, and paid output is watermark-free with commercial rights. The pricing page lists every tier.

Frequently asked questions

What is the difference between image-to-video and reference-to-video?

Image-to-video animates the exact frame you upload, so the composition is fixed. Reference-to-video builds a new scene around elements from your reference images, so the composition changes while the subject stays recognisable. Only the second supports telling a story across different locations.

Which VdoBloom model should I use for reference-to-video?

HappyHorse 1.1 Reference-to-Video is the purpose-built one: 4 to 15 seconds, 720p or 1080p, nine aspect ratios including 21:9 and 9:21. Seedance 2.5 is the alternative if you need longer clips up to 30 seconds or want to attach a reference video, and it is cheaper per second when you do.

How do I stop the character changing between clips?

Use the same reference image for every clip, keep the wardrobe wording identical, describe only the scene in the prompt, and change one variable at a time. If a shot must join tightly to the previous one, use the previous clip’s last frame as the next clip’s first frame.

Can it hold a product consistent rather than a person?

Yes. A bottle, a garment or a piece of packaging works the same way as a face: supply a clean reference and describe the scene around it. The main failure mode is a busy reference, so crop to a single object with plain surroundings where you can.

Is 4K available for reference-to-video?

HappyHorse 1.1 Reference-to-Video runs at 720p and 1080p. For 4K output you need a model with a 4K tier — Seedance 2 reaches 4K, and Kling 3.0 offers a 4K tier as well, though neither is reference-driven in the same way.

→ See HappyHorse 1.1 Reference-to-Video · How it compares to the text-to-video tier · Create a free account