Mario 3D World Selfie Video · Open Skill 공개 스킬
Mario 3D World Selfie Video
Mario 3D World Selfie Video은 Digen의 공개 Open Skill입니다. 열어 보면 이 워크플로가 무엇을 만드는지 알 수 있습니다.
Manhua
4 2 2
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284863/2026-10-19/r7wxw39h.jpg
# Mario 3D World Selfie Video Turn one uploaded reference photo into a short video that looks like a selfie taken inside a bright, toy-like chibi 3D platformer world. The person must stay recognizable, keep their hairstyle, skin tone and outfit accents, and be re-drawn as a chibi character holding the camera at arm's length, while the world behind them is rebuilt from the app's visual language (pastel polka-dot capsule hills, rolling grass, brick blocks, goal pole, coin HUD). There is no structured project state. Everything lives in this conversation plus the canvas assets. Do not assume any character database exists. ## When to use - The user uploads or points at an image and asks for a "Mario 3D world", "chibi world", "mushroom kingdom selfie" image or video. - The user wants to redo the transformation with a different mood, camera distance, outfit, or duration. - The user wants only the still keyframe (skip the script and the video steps). ## Inputs and outputs - Input: one reference photo (user upload, canvas asset, preset asset or library item). Never invent an image URL. - Optional input: a face/character reference for likeness, or a second photo of the same person. - Output: a short landscape (16:9) video, `entity_type='final_videos'`, about 5 seconds. - Intermediate output: a keyframe still, `entity_type='storyboard_images'`, used as the first frame of the video. ## Steps 1. **Confirm the task and collect the photo.** Identify the reference image in the conversation or canvas. If none exists, ask the user to upload one; do not proceed with a guessed subject. If the user asks for a still image only, plan to stop after step 5. 2. **Analyze the reference photo.** Call `recognize_image(image_url=<photo>, prompt=<the identity-descriptor prompt in references/workflow-stages.md>)`. Capture: apparent gender presentation, face shape, hairstyle and hair color, skin tone, eyewear/facial hair, clothing colors and distinct accessories, current pose and camera distance, and background cues worth keeping as a joke or accent. 3. **Read the visual contract.** Call `read_skill_file(name, path="references/style-and-prompt.md")` and `read_skill_file(name, path="references/video-content-constraints.md")`. Do not write a prompt before reading these; they hold the palette, layout, must-keep and must-avoid rules and the reusable prompt templates. 4. **Write the keyframe prompt** by filling the template from `references/style-and-prompt.md` with the identity descriptors from step 2. The prompt must be English and specific to this person; no generic placeholder text may survive. 5. **Generate the keyframe.** `generate_image(prompt=<filled template>, input_urls=[<user photo>], model='i2i.hd.lite', orientation='landscape', resolution='1080P', name='<short title>', entity_type='storyboard_images')` Then run the quality checklist in `references/workflow-stages.md`. If likeness, selfie framing, HUD or the world layout is wrong, regenerate once with a corrective prompt that names the failed items. If it still fails, report honestly and offer to retry with a clearer photo. 6. **Stop here if the user only wanted a still.** Summarize in one or two sentences and call `set_guidance`. 7. **Write this request's video script.** Call `read_skill_file(name, path="references/video-script.md")`, then author a complete script for this specific request (logline, duration, beats with shot size, camera, subject action, environment motion, light, transitions, must-keep, must-avoid). Derive it from the user's wording (mood, energy, duration, camera preference) and from the keyframe. A reused boilerplate script is not acceptable. 8. **Compress the script into the video prompt** using the prompt builder in `references/video-script.md` plus the motion and camera limits in `references/video-content-constraints.md`. One continuous shot, small believable motion, identity locked to the keyframe. 9. **Generate the video.** `generate_video(prompt=<compressed s