English quick take
Generate the still first. Animate one beat: a head turn, a blink, drifting fog, or a slow push-in. Keep the camera language simple and compare the first and last frame.
AI Studio field note 5 min read
Girigo Wish Video: A Five-Second Shot List That Works
A five-second clip is not a mini-movie. It is one shot with one readable change. This guide shows how to plan it, prompt it, and check it before you spend credits.
Quick answer
Field note
One shot, one change
The most common mistake in image-to-video is asking for too much. A costume change, a camera move, a scene transition, and a jump scare in five seconds gives the model no stable frame to hold onto. The face drifts because the model is constantly rebuilding the image. A better shot has one readable change: the head turns, the fog crosses the frame, the reflection arrives late, or the camera pushes in.
Write the shot as a sentence before you write the prompt. "She slowly turns her head toward the camera while fog crosses the garden." That is a shot. "She transforms into a witch and the room catches fire" is a sequence, and a five-second model will struggle to make it coherent.
Field note
Start from a still that already works
Generate or choose a still with a clear face, a strong light source, and enough negative space for the motion. If the still is already ambiguous, the video will be more ambiguous. The still is the identity anchor; the video model only needs to move it, not invent it.
Use the local image editor to test the crop and grade before you animate. A 16:9 or 9:16 frame changes the composition. Decide the final aspect ratio first, then generate the still at that ratio. Cropping a video after generation is possible, but it wastes the model's resolution.
Field note
The Cursed Wish prompt, unpacked
The Cursed Wish recipe gives the model an exact starting composition, preserves the subject's face, costume, hairstyle, and body proportions, and adds one motion beat: a slow head turn, a gentle breeze, drifting fog, and a smooth push-in. It also says no identity change, no costume replacement, no extra characters, no sudden cuts, and no text or watermarks.
That is a lot of constraints for five seconds, but they are all pointing at the same goal: keep the frame stable while one thing moves. If the output is too static, increase the motion clause. If the face drifts, reduce it. Do not change the identity clause and the motion clause at the same time.
Field note
Which model to test
WaveSpeed's Wan image-to-video routes are a sensible budget-first test for subtle motion. Kling routes through Kie or WaveSpeed are useful when you need a more controlled performance beat. fal and Replicate are useful for comparing other hosted models. The exact model names and prices change; check the dashboard before you plan a batch.
Do not choose a model based on a demo reel alone. Run the same still and the same prompt through two routes, then compare the first and last frame. The cheaper route may be enough for fog and a push-in; the more expensive route may be necessary for a head turn.
Field note
The first/last frame check
Export the first frame and the last frame, then compare them side by side. The face should be the same person, the costume should be the same costume, and the light should come from the same direction. If the jawline changes, the eyes move, or the costume changes color, the clip is not usable yet.
Also check the middle frame. Models sometimes produce a stable first and last frame but a strange middle. If the middle is broken, reduce the motion or shorten the clip. A clean three-second shot is better than a broken five-second shot.
Field note
Add the finish in your editor
The model gives you the motion plate. The local video lab gives you a trim, a grade, and a short silent preview. Add sound, titles, and the final compression in your editor. If you need a vertical version, crop from the highest-resolution master rather than regenerating at a different ratio.
The studio page is designed for this handoff: a still-first workflow, three shot recipes, and a local clip lab that does not upload your footage. The model is one part of the pipeline, not the whole product.
Embedded studio
Use the tool while the guide is still open
Open the studio
Girigo Wish Video
The safest first video is a subtle head turn, drifting fog, and a slow push-in. Keep the shot to five seconds, keep one person, and keep the reference image as the visual anchor.
Show the Cursed Wish prompt
Use the provided image as the exact visual reference and starting composition. Create a continuous five-second cinematic Halloween portrait animation. Keep the subject's recognizable face, costume, hairstyle, and body proportions consistent with the reference. The subject slowly turns their head toward the camera with a subtle mysterious expression. A gentle breeze moves loose hair strands and the edges of the clothing. Soft fog drifts through the moonlit garden. The camera performs a slow and smooth push-in. Maintain realistic motion, stable anatomy, and consistent lighting. Do not change the person's identity, replace the costume, add characters, or introduce sudden scene cuts. No text or watermarks.
Model routing
Which model should run this prompt?
| Provider | Best for | Typical route | Docs |
|---|---|---|---|
| WaveSpeedBudget-first | Low-cost image edits and image-to-video drafts | Seedream edit, FLUX Kontext, Wan image-to-video | Docs |
| KieCharacter-first | Nano Banana image edits and Kling video routes | Nano Banana, Kling video | Docs |
| falSpeed-first | Fast hosted queues and strong model coverage | FLUX, Kling, and other hosted media models | Docs |
| SeeAPIRouting layer | One API surface for several model families and fallback routes | Image and video inference endpoints | Docs |
| ReplicateExperiment-first | Broad catalogue for experiments and batch comparisons | Community and first-party image/video models | Docs |
FAQ
Questions this page should answer fast
How long should the clip be?
Five seconds is a good first test. It is long enough for one clear beat and short enough to compare models without a large budget.
Which model is best for image-to-video?
WaveSpeed's Wan routes are a budget-first test for subtle motion. Kling routes are useful for a controlled performance beat. Compare one still and one prompt before a batch.
Why does the face change?
The model is inventing motion without enough identity constraints. Reduce the motion, use a sharper still, shorten the clip, and repeat the identity clause.
Can I edit the video on the page?
Yes. The local clip lab trims a clip, applies a preview grade, and exports a short silent WebM in browsers that support MediaRecorder.
Checked sources