Avatar Camera Angles

Turn one still of a person talking to camera into a set of realistic camera-angle variations of the same recording — same face, same outfit, same room, same light, only the virtual camera moves. Built for vertical reels where the shot cuts every 4–8 seconds while the speaker keeps talking to the lens (the exact pattern seen in real talking-head reels).

This skill is a thin wrapper around the gpt-image-2 skill. gpt-image-2 preserves a reference image's identity at high fidelity; this skill adds the prompt engineering that turns that into a believable camera move instead of a new portrait.

Why it works (validated recipe)

Each prompt is assembled from three parts:

1. Fixed identity/scene block — locks the subject, wardrobe, background props, and lighting so the result reads as another frame of the same video.
2. Camera slot — the single thing that changes, from the move catalog.
3. Framing anchor — stops the model's main failure mode: drifting wider / looser than the source. Most moves keep a tight chest-up frame; pull_out and the negative_space_* moves override the anchor on purpose.

This was validated empirically against a real talking-head frame: identity and the room stay consistent across moves while only the camera position changes.

Setup

Generation runs through gpt-image-2, which owns the Replicate token (shared across the Replicate skills — no separate key needed). This skill only needs Pillow for the optional 9:16 crop:

pip3 install -r ~/.cursor/skills/avatar-camera-angles/scripts/requirements.txt
# gpt-image-2 must be installed too (usually already is):
pip3 install -r ~/.cursor/skills/gpt-image-2/scripts/requirements.txt

Workflow

1. Pick the reference frame. A sharp, front-ish talking-head still of the avatar (e.g. an avatar-frames output). 1 ref is enough; up to 3 helps lock identity.
2. Write the scene profile. Look at the frame and describe four fields — subject, wardrobe, scene, light — in a small JSON file (see examples/lolo_scene.json). This is the most important step: the more accurate it is, the less the scene drifts. You can also pass the fields as --subject/--wardrobe/--scene/--light.
3. Choose camera moves from the catalog (--list).
4. Generate the masters, optionally with --crop916 for the reel-ready 9:16 crop (master 2:3) or --crop169 for a 16:9 YouTube crop (master 3:2).
5. Review & re-roll. Use --count N or re-run a move to pick the best take.
6. Animate each still with lip-sync per shot (the seedance-2 skill or VEED Fabric) and stitch the clips into the reel, cutting between angles every 4–8s.

Quick reference

# Inspect the catalog (no API call)
python3 ~/.cursor/skills/avatar-camera-angles/scripts/generate_angles.py --list

# Preview an assembled prompt without generating
python3 ~/.cursor/skills/avatar-camera-angles/scripts/generate_angles.py \
  --scene-file scene.json --move dutch_tilt --print-prompt

# Generate a few angles (masters in 2:3) + reel-ready 9:16 crops
python3 ~/.cursor/skills/avatar-camera-angles/scripts/generate_angles.py \
  --ref frame_0001.png --scene-file scene.json \
  --move push_in --move low_angle --move three_quarter --move negative_space_left \
  --crop916 -o out/ --slug lolo

# Generate only the empirically validated moves
python3 ~/.cursor/skills/avatar-camera-angles/scripts/generate_angles.py \
  --ref frame_0001.png --scene-file scene.json --validated-only --crop916 -o out/

# 16:9 landscape crops for YouTube (masters in 3:2, cropped to _169.png)
python3 ~/.cursor/skills/avatar-camera-angles/scripts/generate_angles.py \
  --ref frame_0001.png --scene-file scene.json \
  --move push_in --move three_quarter --crop169 -o out/ --slug lolo

Every run prints a JSON object to stdout with a results array (each item has master and, with --crop916/--crop169, reel_916/reel_169). Per-move prompts are saved next to the images as <slug>_<move>.prompt.txt for transparency.

Camera-move catalog

"Validated" = verified to keep identity + scene consistent on a real frame. The others are sound variations; preview/re-roll them as needed.

The pip move (picture-in-picture badge)

When an avatar appears as a small circular badge over a base layer (a broll-web-capture capture, a demo, a terminal), it should be a dedicated shot, not a reused angle: a tight, perfectly centered face close-up with even margin all around (so the circle never clips the face or hair), framed at eye level with no rotation.

• Generate it at 1:1 so the circular crop wastes nothing: --move pip -ar 1:1 (skip --crop916 — the square master is the PiP source).
• Keep it still when you animate it. This still drives a lip-synced clip, but the PiP face must stay put. For the PiP, always lip-sync with avatar-talking-video (p-video-avatar), not seedance-2, keeping the camera locked: pass a --video-prompt like "The person is talking, head still, no camera movement" (no push-in, pull-out, zoom or dolly) so only the face moves. The base layer carries the motion; the avatar is the steady credential anchor.
• Don't burn subtitles into the PiP clip — captions go on the whole reel frame (the reel composer's finish pass), not inside the circle.

Aspect ratio: master 2:3 → 9:16, or master 3:2 → 16:9

gpt-image-2 renders natively only at 1:1 / 3:2 / 2:3. For 9:16 reels this skill generates the master at 2:3 (the cleanest vertical — no canvas padding, full native resolution ~1024×1536) and --crop916 center-crops the sides to a 9:16 reel frame (~864×1536). For 16:9 YouTube pass --crop169: the master defaults to 3:2 and is center-cropped to a 16:9 frame (~1536×864). Both keep native resolution (no upscaling); keep the masters as the archive and feed the crops (_916.png / _169.png) to the reel.

Options

Tips

• Consistency across a reel: always use the same reference frame (and the same scene profile) for every shot, so the avatar doesn't drift between cuts.
• Subtle beats subtle: small camera moves (push_in, low_angle, high_angle) feel like real edits; big ones (profile, hard three_quarter) are punchier but drift more — preview before committing.
• Captions: generate negative_space_* shots for any segment that needs an on-screen headline; the empty side is left clean for text.
• Background variation is fine: the model reconstructs occluded props slightly differently per angle — that actually reads as natural across cuts.
• Faster drafts: use --quality low to scout framings, then re-run the keepers at high.

Related skills

• gpt-image-2 — the underlying image generator.
• avatar-frames — extract clean reference frames.
• seedance-2 — animate each still (lip-sync / motion).
• avatar-video-reel — full reel pipeline.

See REFERENCE.md for the prompt internals and design notes.