Image Expansion Video · Open Skill 공개 스킬
Image Expansion Video
Image Expansion Video은 Digen의 공개 Open Skill입니다. 열어 보면 이 워크플로가 무엇을 만드는지 알 수 있습니다.
Other
0 0 2
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284863/2026-10-19/rkeaqhed.jpg
# Image Expansion Video Turn one uploaded reference image into a short "expansion / inflation" clip: the subject keeps its identity and structure, then visibly swells, rounds out into a glossy balloon-like form, and drifts off the floor — all inside one clean, soft studio look. There is no structured project state. Everything you need lives in this conversation, in the reference image, and in the `references/` files read on demand. ## When to use - The user uploads an image (object, one person, two people, or a multi-person group) and asks for a short video of it expanding, inflating, deforming, ballooning, or floating. - The user asks to redo, lengthen, soften, or restyle an expansion clip you already produced in this conversation. - The user wants a clean studio version of an uploaded subject before animating it. Do not use this Skill for regular image editing with no motion, for full multi-shot editing of existing footage, or for content that sexualizes or degrades real people. ## Steps 1. Restate the brief in one or two lines: the subject in the upload, how many subjects there are, the desired mood, and the target orientation (default `portrait`, 9:16). If the user says nothing about duration, plan a single ~5s clip. Call `write_todos` when the request needs more than one generation pass. 2. `read_skill_file(name="image-expansion-video", path="references/step-details.md")` and follow its input checks and branches. In short: confirm the upload is a real image, count the subjects (single object, single person, two people, and multi-person are all supported; front-facing full-body or upper-body poses give the best result), and decide whether the clip needs multiple passes. If the upload cannot work for this effect, say so plainly instead of generating something unrelated. 3. Analyze the upload with `recognize_image(image_url=<uploaded image>)` when you need concrete facts for the prompt: subject type, clothing and its colors, material and surface, hairstyle, body pose, framing, background, light direction. Only describe what you actually see. Never invent clothing, logos, or background objects that are not in the image. 4. Optional cleanup pass — only when the upload is cropped too tight, blurred, rotated, or has a busy background that would fight the studio look. Call `generate_image` with `model='i2i.hd.lite'`, `input_urls=[<uploaded image>]`, `prompt` built from the start-frame fragment in `references/style-and-prompt.md`, `entity_type='work_in_progress'`, `name='Expansion start frame'`, and `orientation='portrait'`. Use the returned image as the reference for the video. Never re-draw the subject's identity, outfit, or face — this pass only cleans framing and background. Skip this step when the upload is already a clean, front-facing, well-lit image. 5. `read_skill_file(name="image-expansion-video", path="references/style-and-prompt.md")`. From it, assemble the prompt blocks for this request: palette and material (expanded 3D balloon look), subject description taken from step 3, layout and framing, must-keep list, must-avoid list, and the English prompt template. Write the whole video prompt in English. 6. `read_skill_file(name="image-expansion-video", path="references/video-script.md")`. Following that file, write a full English script for **this specific request** — logline, total duration, timecoded beats (input hold → expand & inflate → floating balloon), shot size, camera move, subject action per beat, light and atmosphere, transition, must-keep, must-avoid. Fill it with the user's own words plus the facts from step 3. A generic skeleton is not a finished script; do not paste the template as the deliverable and do not skip this step when the user only wants a still image (in that case stop after step 4/5). 7. `read_skill_file(name="image-expansion-video", path="references/video-content-constraints.md")` and apply its shot, motion, timing, and forbidden-content rules while compressing the script from step