Shared generation policy

<!-- shared-generation-policy -->

Before any paid POST /v1/predictions:

1. Random seed ritual — always first; derive axes via sum-mod.
2. Generation diversity — explicit prompts; rotate ≥2 scenario axes per session.
3. Quality checklists — open output files and judge pass/fail before advancing.

p-video-animate (Pruna)

P-Video-Animate animates a single image using the motion, timing, and camera movement from a given video.

User question this answers: How can I animate this picture with some motion?

Given a reference input image and video, the model generates a new video using (1) the style of the reference image, and (2) preserving the original motion, acting, timing, camera movement, and scene structure of the reference video.

Also on Replicate: prunaai/p-video-animate. Full P-API parameters: p-video-animate model docs · operational guide (Runware host): animating images from video

Shared HTTP patterns: pruna-api.md

p-video-animate vs p-video-replace

Use p-video-replace for in-place identity swap on real clips. Use this skill for motion-transfer showcases, persona variants, and slider before/after demos.

Key features

• Top visual quality
• Most efficient inference
• Speed: 5.24s generation time per 1s video
• Price: $0.03 and $0.06 per 1s of 720p and 1080p video

Key examples

• UGC ad variations — Reuse winning video motions with different creator, customer, or persona images at large scale.
• Viral meme remixes — Turn trending clips into new animated image-based meme formats quickly.
• Movie scene recasting — Animate user-uploaded avatars, selfies, or character images using reusable motion templates.
• Game cinematic variations — Animate user-uploaded avatars, selfies, or character images using reusable motion templates.

Before generating

1. Generation diversity — ritual seed + axis rotation; pass as seed when set.
2. Confirm with the user:

• video URL — motion/audio source (upload .mp4 to /v1/files first)
• image URL — subject to animate (upload jpg/jpeg/png/webp first)
• resolution: 720p or 1080p
• target_fps: original, 24, or 48
• Optional instruction_prompt — how the reference subject should follow the source motion
• Optional save_audio, seed, disable_safety_checker

Run p-video-animate-quality-checklist.md on inputs and outputs.

Batch runs: when several image+video pairs are independent, create all predictions in one parallel async batch, then batch-poll. See parallel-execution.md.

Making motion transfer work

Appearance from image, motion from video. The largest quality factor is how well the reference image matches the first frame of the source video. Ask before every pair:

1. Framing — same body region (head-and-shoulders, medium, full body)?
2. Pose — subject facing and limb position roughly aligned?
3. Visibility — same body parts visible; no extra crop or occlusion the video lacks?

When all three line up, motion transfers cleanly — including secondary cues (camera drift, hair movement, shadow shift). Mismatched pairs lose those.

When pairing fails: head-and-shoulders still + full-body dance → the model does not hallucinate limbs; the head drifts subtly and most choreography is lost. Generate a full-body reference or pick a closer template.

Before generate: repose with p-image-edit when framing or pose is close but not exact. Upscale reference stills for identity detail.

Long clips: ~5 s compute per 1 s of video — split the source and animate segments separately for long templates.

Wrong tool: not for object removal, background replacement, scene rewriting, or frame-exact pixel edits — use p-image-edit, p-video-replace, or inpainting workflows instead.

instruction_prompt — optional; overrides behavior, not who is on screen. Leave blank when the source motion is already right.

Useful (one specific beat on top of source motion):

At the very end of the clip, just after her last gesture, she gives a clear thumbs-up toward the camera. Keep the source motion otherwise.

Less useful (repeats what the image already shows):

A confident woman in a charcoal blazer speaks to the camera in a modern office.

Style variety: photoreal, cartoon, 3D, and mascot stills can share one motion template when framing aligns — the reference image's rendering style carries through.

Runware field map: referenceImages[0] → image, referenceVideos[0] → video, positivePrompt → instruction_prompt, settings.preserveAudio → save_audio.

Mixed-reel context (motion templates, sliders, avatar CTA): animate-beats.md.

Required input

• video (string URL): source RGB video (.mp4); motion and audio source
• image (string URL): reference subject to animate

Common optional fields

• resolution: 720p (default) or 1080p
• target_fps: original (default), 24, or 48
• instruction_prompt (string): extra guidance on how the subject should follow source motion
• save_audio (boolean, default true)
• seed (integer)
• disable_safety_checker (boolean, default false)

Example: upload source assets

curl -X POST "https://api.pruna.ai/v1/files" \
  -H "apikey: ${PRUNA_API_KEY}" \
  -F "content=@/path/to/source-video.mp4"

curl -X POST "https://api.pruna.ai/v1/files" \
  -H "apikey: ${PRUNA_API_KEY}" \
  -F "content=@/path/to/reference-image.png"

Use each response urls.get (or https://api.pruna.ai/v1/files/{id}) in input.video and input.image.

Example: async (recommended)

Omit Try-Sync. Output duration follows the source video.

curl -X POST 'https://api.pruna.ai/v1/predictions' \
  -H 'Content-Type: application/json' \
  -H "apikey: ${PRUNA_API_KEY}" \
  -H 'Model: p-video-animate' \
  -d '{
    "input": {
      "video": "https://api.pruna.ai/v1/files/source-video-abc123",
      "image": "https://api.pruna.ai/v1/files/reference-image-def456",
      "resolution": "720p",
      "target_fps": "original",
      "instruction_prompt": "Animate the reference subject using the motion from the source video."
    }
  }'

Poll and download: pruna-api.md.

Example: sync (single quick test only)

curl -X POST 'https://api.pruna.ai/v1/predictions' \
  -H 'Content-Type: application/json' \
  -H "apikey: ${PRUNA_API_KEY}" \
  -H 'Model: p-video-animate' \
  -H 'Try-Sync: true' \
  -d '{
    "input": {
      "video": "https://api.pruna.ai/v1/files/source-video-abc123",
      "image": "https://api.pruna.ai/v1/files/reference-image-def456"
    }
  }'

Typical next steps

• Generate or edit reference subjects: p-image, p-image-edit
• Replace people in existing footage (not motion transfer): p-video-replace
• Talking-head clips (script-driven, not motion transfer): p-video-avatar
• Pipeline hub: pruna-generative-pipeline
• Multi-scene motion transfer + slider demos: avatar-multi-scene (animate rows)

Related workflow

Animate slider reels: generate_video_comparison.py + avatar-multi-scene animate rows.