Candy Transform Video — Open Skill

Candy Transform Video

Candy Transform Video is a public Open Skill on Digen. Open it to see what this workflow creates.

@eggs

Sports

0 0 0

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284862/2026-10-16/9bu7nbts.jpg

# Candy Transform Video

Turn one uploaded reference image into a short video where the whole frame is restyled as a single candy world: the subject keeps its identity, everything becomes candy material, and the candy bag pours out an endless stream of candies around the subject's feet.

There is no structured project state. One request = one candy keyframe still + one short clip; the stills and the clip land on the canvas automatically.

## When to use

- The user uploads an image and asks for a short video of it "as candy", "candy version", "turn this into candy", etc.
- The user already has a candy still from this skill and wants it animated, re-cut or adjusted.
- The user wants only the candy still (skip the video step if they explicitly say so).

Do not use it for plain image editing that has nothing to do with candy, or for generating an unrelated video from text only.

## Inputs and output

| | |
|---|---|
| Input | One reference image uploaded by the user (asset id / url from the conversation). |
| Output | One short video (candy version of the image), plus the intermediate candy keyframe still. |

The uploaded image is the single source of truth: its subject, pose, expression, count, clothing silhouette, background geometry and camera angle are the "must keep" facts. Never ask the user to re-upload the image, and never invent subjects, props or story beats that are not in the image or in `references/style-and-prompt.md`.

## Steps

1. **Identify the reference image.** Take the image asset from the current conversation (the most recent uploaded image is the source unless the user says otherwise). If its content is not already clear from the conversation, call `recognize_image(image_url=<uploaded image>)` and write down: subject type, number of subjects, pose, expression, clothing silhouette, main props, background, framing and orientation. These details go into the prompt; they are the identity that must survive the restyle.
2. **Read the style conventions before generating.** Call
   `read_skill_file(name='candy-transform-video', path='references/style-and-prompt.md')`
   and follow it. It holds the locked candy style (palette, material, lighting), the layout rules, the must-keep / must-avoid lists, and the copy-ready English image and video prompt templates. Fill those templates with the facts from step 1. Do not paste the reference file back to the user.
3. **Generate the candy keyframe** (image-to-image from the uploaded picture):
   `generate_image(prompt=<image prompt template from references/style-and-prompt.md, filled with the step-1 details>, input_urls=[<uploaded image>], model='i2i.hd.lite', orientation=<see parameters>, resolution='1080P', name='Candy keyframe', entity_type='storyboard_images')`
   The prompt must keep the subject identical and restyle every surface as candy. English prompt only.
4. **Animate the keyframe** (reference-to-video, keeps the restyled subject consistent):
   `generate_video(prompt=<video prompt template from references/style-and-prompt.md>, input_urls=[<candy keyframe asset id from step 3>], model='ref2v', orientation=<same as the keyframe>, name='Candy pour', entity_type='final_videos')`
   The prompt describes one single continuous shot: the seemingly empty candy bag tips over, pours an endless stream of candies that pile up around the subject's feet, and the subject goes from bewildered to happily hugging the bag. English prompt only. Do not pass `model` values other than the channels above, and never a physical model name.
5. **Reply in one or two natural sentences** describing what was produced (candy keyframe + short clip). Do not paste long URLs. If a tool returns `success=false`, say so plainly and do not invent a result or a link.

## Parameters

- `orientation`: match the uploaded image so the framing is not re-cropped — vertical reference → `portrait`, horizontal → `landscape`, square-ish → `square` on the image step (the video model falls back automatically if it doe