Hyperreal Video

hyperreal-video

The opinionated cousin of /generate-video. Always image-to-video (never text-to-video). Defaults to Kling 3.0 because character continuity from the start frame is non-negotiable for believability. Subtle motion vocabulary tuned to match candid energy — not cinematic camera moves that break the illusion.

Providers

• Wavespeed — pull https://wavespeed.ai/docs for the current API surface; model-specific endpoints under https://wavespeed.ai/docs/docs-api/... (e.g. kwai/kling-v3-pro, bytedance/bytedance-seedance-2.0-image-to-video)

When to invoke (vs. /generate-video)

• The video must read as real, not generated
• Character identity must hold from frame to frame — paired with a hyperreal start image
• AI-reveal Reels, candid talking-heads, lifestyle UGC
• The motion is small and natural (talking, looking around, sipping coffee), not cinematic

For multi-cut UGC ads, branded campaigns, or anything text-to-video → use /generate-video instead.

Required input

A start image. Hard requirement. If the user doesn't have one, defer to /hyperreal-image first, then pass the result here. The skill refuses to run text-to-video — that's where character believability dies.

Steps

1. Confirm the start image. Path or URL. If missing, route the user to /hyperreal-image and come back.
2. Capture the motion. What should the character do? Talk to camera? Look down at phone? Sip coffee? Adjust hair? Keep it small and natural — one or two beats, not a multi-shot scene.
3. Capture the speech (optional). If they're talking, get the dialogue. Note: Kling has no native audio — pair with /clone-voice
   • /swap-voice after generation if dialogue is needed.
4. Refine the motion prompt using the template and vocabulary below. Read it back; iterate before any provider call.
5. Pick the model. Kling 3.0 default (best character continuity from a start frame). Seedance 2.0 only if the user explicitly wants native audio and accepts identity-drift risk.
6. Set the params. Aspect ratio (matches the start image — 9:16 default), duration (5–8s default for Kling believability sweet spot), resolution.
7. Generate via Wavespeed image-to-video. Download the MP4 to CWD/hyperreal-video-<timestamp>/clip.mp4.
8. Show the user the result. If dialogue was specified, prompt them to run /clone-voice + /swap-voice for audio.

The template (load-bearing)

Subtle natural motion: [character does WHAT — single beat], [camera — handheld micro-movement / static], [scene reads as candid phone capture]. Character identity locked from start frame. [duration] seconds. [aspect ratio].

Motion vocabulary (use these — they preserve believability)

• subtle natural motion
• small head turn
• slight head tilt
• looking down at phone
• talking to camera
• slight smile
• breathing
• adjusting hair
• sipping [coffee / drink]
• handheld micro-movement
• static frame, character motion only

Avoid (kills believability)

• cinematic camera move, dolly, tracking shot, crane, drone
• multi-shot, cut to, then cuts to
• dramatic lighting change
• hair flip, dramatic gesture
• Anything implying multiple shots, scene changes, or styled action

Negative prompt

CGI, 3D render, video game character, AI generated look, plastic skin, smooth plastic skin, beauty filter, multi-shot, cut, scene change, multiple lighting setups, dramatic camera move, dolly zoom, crane shot, tracking shot, cinematic, stylized motion, exaggerated facial expression, uncanny valley

Pick the model

Defaults

Input

• Start image (path or URL — required)
• Motion description (what the character does)
• (Optional) Dialogue, if speech is needed
• Duration
• Aspect ratio (defaults to start image's ratio)

Output

• MP4 at CWD/hyperreal-video-<timestamp>/clip.mp4
• Wavespeed-hosted URL
• Final motion prompt saved alongside
• If dialogue was specified, a follow-up note pointing to /clone-voice + /swap-voice

Common failure modes

• Character drifts after frame 1 — start image was too dynamic (extreme angle, motion blur). Use a clean medium close-up start frame.
• Cinematic camera move appeared — prompt mentioned "shot," "scene," or any cinematography term. Strip to motion-only vocabulary.
• Robotic / unnatural motion — described too many beats. Cut to one or two micro-actions.
• Lips out of sync with intended dialogue — Kling has no audio model; lips will move generically. Run /swap-voice to align dialogue precisely.

How it pairs

The natural pipeline:

1. /hyperreal-image → 2K PNG of believable character in setting
2. /hyperreal-video → MP4 with subtle motion, character identity locked from the PNG
3. /clone-voice → voice ID
4. /swap-voice → MP4 with the right voice talking
5. /add-captions → final post-ready MP4