Funny play game comedy with g.f Scene Reel · skill Open Skill công khai

Funny play game comedy with g.f Scene Reel

Funny play game comedy with g.f Scene Reel là một Open Skill công khai trên Digen. Mở ra để xem workflow này tạo gì.

@toopchat

Other

2 1 2

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4239354/2026-10-21/39yyfkyw.jpg

---

allowed-tools:
  - generate_image
  - generate_video
  - recognize_image
  - recognize_video
  - probe_media
  - mix_clips
  - generate_music
  - create_document
---

# Timeline Scene Reel

Turns a natural-language, time-stamped scene brief into a single-take photorealistic vertical (9:16) reel, with character faces locked to reference photos.

## When to use
- The user gives a beat-by-beat timeline ("0–2 s … 2–4 s …") and wants one continuous shot.
- The user wants the same recurring characters (from reference photos) to look identical across many different scenes.
- The user wants to redo/restyle/re-time an existing reel of this kind.

## Steps

1. **Parse the brief into beats.** Extract: total duration, per-beat time windows, who is on camera, the action, the camera style, the look (lighting, room, lens feel), the audio intent, and any negative prompt the user gave. Keep the user's own wording for actions — it is the ground truth for the video prompt. If a beat window is missing, infer a reasonable split and say so in your reply.

2. **Resolve the cast (face consistency).**
   - For every recurring person, find their reference photo: an asset the user uploaded/attached this turn, an asset already in this session, or a bound preset asset (see `references/cast-library.md` for the bound keys).
   - If a person has **no** reference photo, generate one first with `generate_image` (`model='i2i'` if you have a similar face to anchor on, else `model='t2i'`), a clean chest-up photorealistic portrait, neutral background, natural light, matching the age/clothing described in the brief. Then use that generated portrait as their reference. Pass `name` = "<Person> reference portrait", `entity_type` = `characters`.
   - Optional but useful: `recognize_image` on an unfamiliar reference photo to confirm age, headscarf, clothing, and setting before writing the prompt, so the prompt matches the actual face.

3. **Write the video prompt** using the structure in `references/prompt-recipe.md`. One paragraph per beat, in order, with explicit time windows; a fixed global look line (photorealistic, cinematic, camera model, lighting, room); explicit "same face, same clothing, consistent identity throughout" reinforcement; the audio line; and the negative line. Prompts are written in English.

4. **Generate the clip.**
   - `generate_video` with `model='ref2v'` (reference-to-video) and `input_urls` = the reference photos of every on-camera character, in a fixed order (man first, woman second, etc. — keep the same order across takes).
   - `orientation='portrait'`, `resolution='1080P'`, `duration` = the brief's total seconds (clamped automatically if the channel caps it).
   - `name` = "<Scene title>", `entity_type` = `final_videos`.
   - For a shot that must start from one exact composition, you may instead prepare a first frame with `generate_image` and use `model='i2v'`; prefer `ref2v` when the priority is face consistency.

5. **Add sound only if asked.** If the brief wants background music, generate it with `generate_music` (`model='music'`, `entity_type` = `music`) and mix it under the clip with `mix_clips` (clip video + music audio, `name` = "<Scene title> final", `entity_type` = `final_videos`, `orientation='portrait'`). Keep natural in-scene sound in the generated clip; do not add dialogue or voice-over unless the user asked for speech.

6. **QC the result.** Run `recognize_video` on the finished clip with a prompt that checks exactly the beats from step 1: does each beat happen, is the camera fixed, do the faces match the reference, is anyone deformed or duplicated. If a beat is missing or a face drifts, re-run step 3–4 with a corrected prompt (add the missing beat explicitly, restate the reference anchor at the top) rather than patching after the fact. `probe_media` if you need the real duration before trimming.

7. **Report.** Summarize the beat list, the reference photos used, the channel and settings, and where the clip landed. Ment