WildClips v4.4 prompt guide

How to write prompts for video from frames and references: structure, shots, timestamps, and image alignment: 0:

Choose a mode

WildClips v4.4 generates video with audio in two modes:

Frames — a first frame, a last frame, or both. The video starts from the first frame and arrives at the last one.
References — up to 4 images. The model transfers the characters' appearance, clothing, objects, setting, or style from them, but the video frame does not have to repeat the reference.

Frames and references cannot be combined in one generation.

Prompt structure for frames

The prompt consists of three fields. Keep the field names in English as shown; write their content in any language.

integrated_multimodal_description: [Shot 1] ...

overall_soundscape: ...

non_diegetic_music: ...

integrated_multimodal_description — the visuals: style, composition, characters, actions, camera movement, and dialogue in the order they happen.
overall_soundscape — ambient and action sounds: footsteps, wind, rain, breathing, city noise.
non_diegetic_music — background score that only the audience hears. Write N/A if no music is needed.

The first shot is [Shot 1] without a timestamp. Each later shot starts with its label and the cut time in the final video: [Shot 2] At 00:03.000, [Shot 3] At 00:06.500. Times must increase and fit within the selected duration.

A short clip usually needs one shot with one clear action. Add a cut only when it shows something new. In the first shot, establish the style, composition, the characters' appearance and position, the setting, and the lighting, then describe the action in order.

Align frames with the timeline

Before the fields, add a line that tells the model where the uploaded frame belongs in the video. Replace 10.00 with the selected duration.

First frame only:

For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced.

Last frame only (N is the number of the final shot):

How the reference pictures align with the target video — <Picture 1> (from [Shot N]) aligns with the 10.00-second mark of the target video.

First and last frames:

How the reference pictures align with the target video — Picture 1 (from Shot 1) aligns with the 0.00-second mark of the target video; Picture 2 (from Shot N) aligns with the 10.00-second mark of the target video.

With a first frame, describe what starts to happen rather than the picture itself. With a last frame, describe the path that gradually leads to it. With both frames, describe the transition between them instead of two separate still images.

First-frame example:

For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced.

integrated_multimodal_description: [Shot 1] Starting from the still portrait, the woman's eyes slowly open. Her short dark hair drifts gently in a soft breeze. Behind her the cherry blossom petals begin to fall, drifting past her face. [Shot 2] At 00:04.000 she turns her head slightly left, catching one petal on her fingertip, the background shifting from blurred garden into sharp focus revealing traditional lanterns hanging among the branches.

overall_soundscape: Gentle spring breeze, distant birdsong, subtle rustling leaves and falling petals.

non_diegetic_music: Delicate koto and shakuhachi melody begins softly after 00:01.000, swelling warmly by the final second.

Prompt structure for references

The prompt consists of six sections in this order:

subject_definitions:
<Subject 1> ...

summary:
[reference generation] ...

retention_analysis:
<Subject 1> (appears in [Shot 1]): fully_preserved — ...

detailed_description:
...

overall_soundscape:
...

non_diegetic_music:
...

subject_definitions — each character, object, or place as <Subject N>: which image it comes from and which features matter. If an image defines a specific frame composition, refer to the image itself.
summary — one short paragraph that starts with [reference generation].
retention_analysis — one line per label stating how closely to keep the reference: fully_preserved — completely, partially_preserved — partly, attribute_transfer — transfer selected features only, weak_reference — use as loose guidance.
detailed_description — the style in one or two sentences, then [Shot N] shots with timestamps.
overall_soundscape and non_diegetic_music — same as in frame mode.

Use each label with the same meaning in every section.

Reference example:

subject_definitions:
<Subject 1> is the woman in @Image1: short red hair, freckles, green eyes, an olive jacket over a white T-shirt.

summary:
[reference generation] A ten-second 16:9 cinematic video in which <Subject 1> finds an old photograph in a bookshop and smiles.

retention_analysis:
<Subject 1> (appears in [Shot 1]-[Shot 2]): fully_preserved — keep her face, hairstyle, freckles, and clothing in every shot.

detailed_description:
Live-action footage with warm evening light and a shallow depth of field.

[Shot 1] Medium shot. <Subject 1> walks along the bookshelves, runs her fingers over the spines, and pulls out a worn volume; an old photograph slips out of it. The camera slowly tracks behind her.

[Shot 2] At 00:05.000 close-up. She picks up the photograph, studies it, and laughs quietly; the camera pushes in toward her face with small amplitude.

overall_soundscape:
Quiet shop ambience, rustling pages, footsteps on a wooden floor, soft laughter.

non_diegetic_music:
Calm piano that gently swells in the second shot.

General tips

Describe only what can be seen and heard: gaze, facial expression, posture, gestures, object movement.
Name the camera movement: push in, pull out, pan, tilt, tracking shot, arc shot, static shot. Add amplitude and speed when they matter.
Put on-screen text and exact dialogue in quotation marks.
Keep appearance, clothing, lighting, and character positions consistent between shots.
Remove contradictory instructions and vague phrases such as "make it beautiful".

Use prompt enhancement

Use Enhance prompt when the initial idea is short. The enhancer builds the required structure and alignment lines from the uploaded frames or references and the selected duration, while preserving your scene and the prompt language.