Use a first frame when the opening composition and identity matter most. Add a last frame when the shot must land on a defined final composition. Use multiple references when different files need to supply identity, product views, environment, camera movement, or audio rhythm. Do not add every control automatically; each one narrows the model's room to solve motion.
Oakgen exposes different fields by model. Seedance 2.0 image-to-video currently accepts an opening image and optional last image, while Seedance 2.0 text-to-video accepts optional image, video, and audio references. Check the selected workflow rather than assuming controls transfer across models.
Decision table
| Job | First frame | Last frame | Multiple references | Main risk |
|---|---|---|---|---|
| Animate a portrait reaction | Yes | Usually no | Optional profile | Face drift during large turn |
| Product beauty push-in | Yes | Optional | Product angles if orbiting | Geometry invention |
| Before/after composition | Yes | Yes | Usually no | Impossible interpolation |
| Character in a new setting | Optional | No | Identity + environment | References competing for style |
| Copy a camera pace | Optional | No | Add short video reference | Source scene leaks into output |
| End on approved packshot | Yes | Yes | Product angles optional | Final frame mismatches opening |
| Multi-shot sequence | Optional | Optional | Identity, setting, motion, audio | Too many constraints for duration |
| Music-led visual | Optional | No | Audio + identity/style | Rhythm reference mistaken for content |
Choose the Smallest Control Set
Open Oakgen, pick the workflow that exposes the control you need, and test one short action before adding more references.
First-frame image-to-video
The opening image provides composition, subject identity, lighting, and a starting pose. It works best when movement remains close to what the image proves.
Use it for slow push-ins, a small expression change, fabric movement, a product lid opening, or a short environmental effect. Prepare the image with the product video prompt guide.
Avoid a wide orbit from one flat product photo. The model has no evidence for the back.
Last-frame control
A last frame gives the shot a destination. It can help with a product closing, a match transition, a before/after composition, or a planned final hold.
Check four things before adding it:
- the same subject appears in both frames;
- scale and perspective can change through a believable camera or subject move;
- lighting belongs to the same sequence, unless the transition itself explains the change;
- the duration gives the model enough time to travel between states.
Two beautiful frames can still form an impossible shot. A seated front portrait and a distant rear running pose may need a cut, not interpolation.
Multi-reference control
Several references help when one source cannot define the job. Assign ownership:
Image 1: character face, hair, and wardrobe.
Image 2: product geometry and material.
Image 3: market environment and lighting.
Video 1: side-tracking camera height and pace only.
Audio 1: percussion tempo only.
Conflicts become visible on paper. If image 1 shows warm daylight and image 3 shows a rainy night, state that the environment owns lighting.
Control budget
Treat every added demand as spending a limited control budget:
| Demand | Low | Medium | High |
|---|---|---|---|
| Identity change | Small expression | Head turn | New angle plus wardrobe |
| Product motion | Static push-in | Lid or hinge | Full hand interaction and orbit |
| Camera | Locked | One push or pan | Compound orbit, crane, and zoom |
| Environment | Stable | Weather movement | Location and lighting transformation |
| Timeline | One action | Two beats | Several cuts and interactions |
| References | One clear source | Two assigned roles | Many overlapping roles |
If several rows are high, split the shot. Models can perform difficult work, but a production plan should not depend on every difficult condition succeeding at once.
Three setup examples
Product page loop
Use one clean first frame, a slow push-in, no last frame, and no reference pack. The job is scale and texture, not reinvention.
Packshot transition
Use a lifestyle first frame and a studio last frame only if the product has the same proportions and view. Prompt the environmental transition explicitly and reserve enough duration.
Character-led campaign shot
Use a character identity image, product image, environment image, and one short camera reference. Do not add a last frame until the sequence itself holds, because the reference pack already gives the model several obligations.
How Seedance 2.0 maps to these choices
ByteDance describes Seedance 2.0 as accepting text, image, video, and audio references upstream. Oakgen's current implementation separates the controls:
- Seedance text-to-video: prompt plus optional image, video, and audio references;
- Seedance image-to-video: required opening image plus optional last image.
Oakgen currently exposes no audio-generation toggle for these forms, though reference audio is available in text-to-video. Provider requests can use endpoint defaults, so verify sound in the returned file.
Failure patterns
The model ignores one reference. Two files may own the same job. Remove one or state a strict role.
The middle morphs between good first and last frames. The endpoints are too far apart for the duration. Create an intermediate shot or use a cut.
The source scene leaks from a motion reference. State that the video supplies only camera path and pace; reduce visual similarity between the reference subject and final subject if possible.
Identity gets weaker with more images. The identity set contains conflicting faces, lighting, ages, or hairstyles. Curate it rather than expanding it.
FAQ
Is a last frame always better?
No. It can force an awkward path between incompatible compositions. Use it only when the destination matters and the transition is physically plausible.
How many reference images should I use?
Use the smallest set that defines the views and roles your shot needs. One strong source can beat six loosely related ones.
Can a video reference control camera movement?
Seedance 2.0 accepts video references, and ByteDance says references can supply camera movement and motion cues. State that role directly and do not assume perfect replication.
Does reference audio mean the output contains audio?
Not necessarily. Input and output are separate capabilities. Oakgen accepts Seedance audio references but exposes no audio-generation toggle, so inspect the returned file and plan a separate sound pass when delivery audio must be controlled.
Should I use first and last frames for a product rotation?
Only if both frames depict the exact same product and the angle change is modest. Multiple product angles may be better evidence for a larger reveal.
What if my model does not expose these controls?
Choose another model or simplify the shot. Prompt language cannot create an input field that the host does not provide.
Continue the workflow
- Prepare the source with the product video prompt guide.
- Repair character and product drift.
- Read the Seedance 2 control ledger.
- Build a three-beat Seedance sequence.
Match the Control to the Shot
Choose a first-frame, last-frame, or reference-led Oakgen workflow based on the one detail that must remain exact.
