Text-to-Video (T2V)
You describe a scene in words; the model invents the rest: subject, environment, motion, camera, and (on most 2026 models) audio. No input image anchors the result, so you trade control for creative freedom.
- Use it for: concepting, B-roll, abstract or fantastical scenes, anything where you don't already have a subject image.
- Strength: maximum imagination, fastest to start.
- Weakness: least control over exact appearance; the same prompt yields different faces/looks each run.
Pro move: If you need a specific look, generate a still image first (T2I), then feed it into I2V. You get the imagination of text plus the lock-in of an image.
Image-to-Video (I2V)
You supply a starting image (and sometimes an ending frame); the model animates it. Appearance, composition, lighting, and identity are largely "locked" by the image, so your prompt mostly controls motion and camera.
- Use it for: bringing a specific character/product/photo to life, consistent looks, controlled results, most mature content workflows.
- Strength: highest fidelity to a known subject; far fewer "lottery" re-rolls.
- Weakness: input quality caps output quality. Blurry image in, blurry video out.
First/last frame: Kling 3, Seedance 2.0, and Wan 2.7 support setting both a start and end frame. This is the most reliable way to choreograph a precise transformation (pose A → pose B).
Reference-to-Video (R2V)
The most powerful and most misunderstood mode. Instead of one start frame, you feed the model a library of references (character images, a motion clip, a style board, an audio track) and you tell it, in your prompt, what to take from each. The model extracts those elements and builds a new video.
This is where Seedance 2.0 and Wan 2.7 shine. A single generation can combine, for example: the character from Image 1, the camera move from Video 1, the lighting style from Image 2, and the voice timbre from Audio 1.
- Use it for: character consistency across shots, motion transfer (copy a dance/camera move onto a new subject), style borrowing, voice cloning, multi-character scenes.
- Strength: director-level control; combine elements no single image could capture.
- Weakness: steeper learning curve. You must say which reference governs which dimension, or the model guesses.
R2V sub-tasks (Seedance terminology)
ByteDance's own guide splits reference workflows into three task types, a useful mental model for any model:
| Task | What it does | Prompt pattern |
|---|---|---|
| Reference | Extract elements (subject, style, motion, sound) to make a new video. | "Refer to the [action/style/sound] in @Video1 to generate…" |
| Edit | Modify part of an existing video; everything unmentioned stays the same. | "Strictly edit @Video1, changing its [original feature] to [new feature]…" |
| Extend | Continue a clip forward (or backward) with consistent identity and style. | "Extend @Video1, generate…" |
When editing or extending, reference the clip directly (e.g. @Video1) rather than saying "Reference Video 1," which the model can misread as a new referencing task.
Watch: Reference-to-Video in Venice Studio
Full library: Video Guides.