Street Bend I2I Video · skill pública de Open Skill
Street Bend I2I Video
Street Bend I2I Video es una Open Skill pública en Digen. Ábrela para ver qué crea este flujo.
Custom
0 0 0
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284863/2026-10-20/ajnunggk.jpg
--- allowed-tools: - recognize_image - generate_image - generate_video --- # Street Bend I2I Video Turns one uploaded image into a cinematic short video: the subject from the photo is re-staged at a bend in a coastal street wrapped in cold fog, then animated with a slow forward glide along the curve. Pipeline is always **reference analysis -> restyled keyframe (image-to-image) -> written shot script -> video**. There is no structured project state; the keyframe and the clip live on the canvas, so keep track of which canvas asset is the current keyframe and which is the current clip in your replies. ## When to use - The user uploads a picture (person, object, animal, scene) and asks for "a video like this", "make it move", "turn this photo into a clip". - The user asks for the foggy street-bend look, a cinematic restyle of their photo, or a moody atmospheric loop. - The user only wants to change part of an existing result (the keyframe look, the camera move, the mood, the length). In that case redo only that stage - do not restart the whole pipeline. - The user asks for still images only: run stages 1-3 and skip the script and the video. Do not use for: full video editing or montage of existing footage, subtitle work, or multi-scene narratives with hard cuts. ## Inputs and outputs | Item | Rule | |---|---| | Reference image | Required. It defines identity, wardrobe, pose family and framing that must survive the restyle. | | Text from the user | Optional but authoritative. It can override wardrobe, weather, time of day, mood, orientation and duration. | | Output | One vertical-by-default short video (about 5 seconds), single continuous shot. | | Intermediate | One restyled keyframe image, shown to the user so they can approve the look before paying for video. | How each input is used in prompts: the image feeds `input_urls` (image-to-image) and supplies the subject block; the user's words fill the setting, weather, mood, camera and duration slots. When the two conflict, the user's words win for everything except the subject's identity, which the image always wins. ## Steps 1. **Intake.** Confirm you have a reference image. If not, ask for one and stop. Collect any text the user gave about subject, wardrobe, place, weather, mood, orientation, duration. Create a todo list with `write_todos` (analyse -> keyframe -> script -> video -> deliver). 2. **Analyse the reference.** Call `recognize_image` with `image_url=<reference>` and a prompt asking for: subject and visible identity cues, wardrobe and materials, pose, framing and camera height, dominant light direction, palette, and any background elements worth keeping. Write the findings down in your reply in a few short lines so the user can correct you. 3. **Read the style reference.** Call `read_skill_file(name="street-bend-i2i-video", path="references/style-and-prompt.md")`. Follow its Style / Subject / Layout / Must keep / Must avoid rules and fill its keyframe prompt template with the analysis from step 2 plus the user's text. Do not improvise a look that contradicts that file. 4. **Read the stage rules.** Call `read_skill_file(name="street-bend-i2i-video", path="references/workflow-stages.md")` and follow the stage checklist, the branch for "use my photo unchanged", and the retry rules. 5. **Build the keyframe.** Call `generate_image` with the filled prompt, `input_urls=[reference]`, `model='i2i.hd.lite'`, `orientation` (`portrait` by default, `landscape` on request), optional `resolution`, a short `name` (e.g. "Street bend keyframe"), and `entity_type='storyboard_images'`. Present the keyframe and ask for approval only if the user asked to review before video; otherwise continue. 6. **Write the shot script.** Call `read_skill_file(name="street-bend-i2i-video", path="references/video-script.md")`. Then write a real, specific script for *this* request following that structure - logline, duration, beat-by-beat framing, camera path, subject action, light and