First and last frames describe where a video begins and ends. Reference images and clips describe what a subject looks like or how it should move. Choose boundary frames for a controlled transition and reference-to-video for separate identity, motion or audio guidance. Some operations can combine both.
This distinction is useful when choosing among Kling 3, WAN 3 and MiniMax H3 workflows. It also prevents a common integration mistake: treating every image input as if it had the same meaning.
Match the input to the constraint
| What must be controlled? | Input to look for | What it does not establish by itself |
|---|---|---|
| Opening composition | First image | Exact identity throughout the whole clip |
| Closing composition | Last image | The motion path between the endpoints |
| Character or product identity | Subject reference or element | A specific camera endpoint |
| Movement or camera style | Reference video or motion-control operation | Correct product details |
| A specific midpoint | Timed middle-image control | General availability across every model |
| Rhythm or sound context | Reference-audio input | Exact reproduction of the recording |
The prompt connects those constraints. It should describe the intended action and camera movement without assigning conflicting roles to the same reference.
Which reviewed operations expose boundary frames?
On 19 October 2026, the reviewed non-Turbo Kling 3 image-to-video operations exposed first and last images. WAN 3 and Raywake's WAN 3 Prime image-to-video operations also exposed opening and closing images, with different names for the first-image field.
MiniMax H3 Max Reference offered first and last images alongside multimodal references. It additionally exposed a timed middle-image control under specific resolution and input conditions. The reviewed Kling Turbo image-to-video contract did not expose the same final-frame control.
Sources: Kling Standard, WAN 3, WAN Prime, H3 Max Reference.
When should you use first and last frames?
Use them when the start and finish are part of the deliverable: a product moves into an approved hero shot, a room changes lighting, or a camera transitions between two planned compositions.
Prepare the pair carefully. Keep product geometry, label, lighting logic and intended aspect ratio compatible. If the first image shows one bottle and the last shows a redesigned bottle, the model has contradictory identity information. A smooth-looking transition can still be wrong for the product.
Describe one plausible path between the images. A short clip with several reversals, scene changes and transformations can be harder to control than a single camera move. Review the entire middle section, because boundary frames alone do not establish stable interpolation.
When should you use reference-to-video?
Use it when different assets supply different information. A character image can establish appearance, a movement clip can establish choreography and an audio segment can establish a timing cue. Their roles need to be explicit.
For a recurring mascot, keep the identity reference consistent across shots while changing the environment and action. For a product, use an approved packshot as the authority and avoid references containing competing labels or shapes.
Available Raywake candidates include WAN 3 Prime Reference, MiniMax H3 Reference and H3 Max Reference. Their input counts, duration limits and pricing behavior differ; inspect the selected operation instead of building one universal request body.
Can both approaches be combined?
Yes, when the chosen operation explicitly supports them. H3 Max Reference is one reviewed example. Its frame controls and reference inputs can describe different parts of the same brief, but compatibility conditions still apply.
Keep the combined request as simple as possible. If an identity reference conflicts with a boundary image, decide which asset is authoritative before generation. Adding more files does not resolve contradictory instructions automatically.
The H3 comparison explains the middle-frame restrictions. The WAN comparison shows why field mappings must be reviewed when changing operations.
A practical five-second evaluation
Choose a single approved product and a simple move. Define four checks: unchanged geometry, readable label, requested motion and an acceptable ending. Use a fixed duration and record the exact input files and settings.
Quote the selected request in Raywake before generation. After completion, inspect first, middle and last sections, then watch the clip continuously. A frame-by-frame check and normal-speed playback reveal different problems.
Record the settled cost and the reason for any retry. This makes the next test a deliberate change rather than a search for a lucky result. For a longer campaign, reuse the accepted brief structure and change only the part needed for the next shot.
Common questions
Is a subject reference an exact first frame?
No. It can guide identity without fixing the opening composition. Use a first-image control when the opening itself matters.
Does a final image guarantee a perfect transition?
No. It supplies a boundary constraint. Motion, continuity and product identity still need review.
Can the same seed make two models comparable?
A seed is not a universal visual identity across models. Match the actual brief and shared controls, and record differences you cannot match.
Choose the right video inputs in RaywakeReviewed on 19 October 2026. The cover is an editorial illustration generated with Nano Banana 2; it is not a test result from the models discussed.