chase-scene-video · публичный навык Open Skill

chase-scene-video

chase-scene-video — публичный Open Skill на Digen. Откройте, чтобы увидеть, что создаёт этот процесс.

@socleva24

ShortDrama

6 0 3

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/3770114/2026-09-18/wyed35sh.jpg

# Chase Scene Video

## When to use
- The user provides a character image, a car image, a house image, and a chase/action description and wants a video of the scene.
- The user wants to redo a single shot, add more beats, or combine the shots into one clip.

## Inputs (flexible — the user supplies these each time)
- character image (the protagonist, e.g. the woman)
- car image (the protagonist's car)
- house image (the starting location)
- chase/action description (what happens, e.g. "she leaves the house, hops into her car, drives on a curvy hilltop road, and a dark car suddenly pursues her")

## Steps
1. First call `write_todos` to list the beats you'll produce, then update it as you go.
2. Parse the description into **beats** — each beat becomes one separate shot. Example beats from the sample description:
   - Beat 1: woman leaves the house
   - Beat 2: woman gets into her car in the driveway
   - Beat 3: woman drives on the curvy hilltop road
   - Beat 4: a dark car suddenly appears behind and pursues her
3. For each beat, generate one video shot with `generate_video`:
   - `model='ref2v'` (reference-to-video, keeps the character/car/house consistent)
   - `input_urls` = the reference images relevant to that beat (character, car, house — include only the ones actually in the shot)
   - `prompt` = an English description of that single beat, keeping the character/car/house consistent with the references
   - `orientation` = the user's choice (default `landscape` for cinematic; `portrait` for vertical short-form)
   - `resolution` = `720P` by default
   - `name` = e.g. "Shot 1 - leaves house"
   - `entity_type` = `chase_shots`
4. Present each shot to the user. Offer to redo any shot, add more beats, or stitch them in order with `mix_clips` (pass the shots as `clips[].video.url` in sequence, `entity_type='final_videos'`).

## Notes
- Video generation does **not** support NSFW/adult content on any channel — decline such requests politely.
- Report failures honestly; never fabricate URLs or claim a shot succeeded when it didn't.
- Each beat's shot is a separate asset and is visible on the canvas (intermediate products are not hidden).
- When a stage completes, call `set_guidance` to suggest next actions (e.g. redo a shot, add a beat, stitch the shots).
- **Storyboard trick for stubborn compositions**: if a shot's composition keeps coming out wrong no matter how you re-prompt `ref2v` (a common case is the chase direction getting inverted — the pursuing car ends up in front of the chased car instead of behind it), don't keep re-prompting. Instead: (1) generate a clean storyboard image with `generate_image` (`t2i`, or `i2i` if you want to reuse a reference) that explicitly locks in the correct composition (e.g. "camera behind the dark car, white car ahead, dark car pursuing from behind"); (2) verify it with `recognize_image` before animating; (3) animate it with `generate_video` using `model='i2v'` and the storyboard image as the start frame. This reliably fixes direction/composition problems that `ref2v` keeps inverting.