Outfit Transfer & Showcase · Open Skill 공개 스킬

Outfit Transfer & Showcase

Outfit Transfer & Showcase은 Digen의 공개 Open Skill입니다. 열어 보면 이 워크플로가 무엇을 만드는지 알 수 있습니다.

@lukasolxxx

TryOn

6 0 7

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/2175537/2026-10-18/cvnhkdkw.png

---

allowed-tools:
  - recognize_image
  - generate_image
  - generate_video
  - generate_music
  - extract_frame
  - mix_clips
  - probe_media
---

# Outfit Transfer & Showcase

## When to use
- The user provides two reference images (a model/character and an outfit) and wants the model to wear the outfit, then a three-shot video showcasing it.
- Outfit transfer / clothing swap / "dress this model in this outfit" / outfit showcase video.
- Also when they want to redo or tweak any stage (transfer image, a specific shot, or the final mix).

## Preset assets
- `skill_cover`: the skill's cover image (fashion showcase cover with the "OUTFIT TRANSFER & SHOWCASE" title). Use it as the representative cover/thumbnail for this skill.

## Inputs
- INPUT A — MODEL REFERENCE: defines identity (face, hair, skin tone, body proportions, age). Always the source of truth for WHO the character is.
- INPUT B — OUTFIT REFERENCE: defines the clothing (design, color, pattern, material, construction, accessories). Always the source of truth for WHAT the character wears.
- If the user gives two images without labels, treat the first as model, the second as outfit.

## Reference files
- `references/prompts.md`: main video prompt template, negative prompt, and per-shot prompts. Read it before generating any video.

## Model notes
- Preferred video-generation backends: Kling 2.1, Runway Gen-3, Luma, Wan, Hunyuan, Real Motion.
- These are informational notes on the underlying models. Generation calls still use stable channels (e.g. `ref2v` for the shots), which route automatically to the best available backend — do not pass these names as the `model=` parameter, since `generate_video` rejects concrete model names.
- Image quality: generate the outfit-transfer image at 2K (resolution='2K') for maximum detail. Note that portrait video output is capped at 480p by the platform regardless of input resolution.

## Camera direction
Use stable framing and smooth, controlled movement. Avoid excessive zoom, abrupt cuts, sudden rotation, uncontrolled tracking, extreme perspective distortion, excessive motion blur, and unnecessary cropping. Keep the complete outfit visible and coherent framing between shots.
- Shot 1 (establishing): full-body, eye-level, centered, static camera with minimal movement — show the whole outfit and silhouette.
- Shot 2 (side profile): full-body or medium framing, camera slightly angled to follow the model's side/three-quarter profile; controlled movement; a close-up may be used for visible garment details.
- Shot 3 (final presentation): full-body or three-quarter framing, stable, outfit clearly visible; medium/close-up allowed for final detail.

## Frame lock & extraction (important)
- LOCK: the final frame of each scene is the reference for the next shot. Shot 1's final frame becomes the starting visual reference for Shot 2; Shot 2's final frame becomes the reference for Shot 3. Keep the original model and outfit references active throughout as identity/clothing constraints.
- Extract the final frame at 1K quality: call extract_frame with width=1080 so the reference frame is high-resolution.
- Validate each extracted frame (recognize_image) before reusing it; if it is blurred/corrupted/distorted, extract the nearest clean earlier frame instead. Never propagate a corrupted frame into the next shot.

## Steps
1. First call write_todos to lay out the pipeline, then update the whole list as you progress.
2. As soon as the model reference is provided, immediately read/analyze it with recognize_image and lock it as the identity source (face, hair, skin tone, body shape/proportions) before any other step. Then analyze the outfit reference the same way and lock it as the clothing source.
3. Outfit transfer — generate_image model='i2i', input_urls=[model_ref, outfit_ref], orientation='portrait' (9:16; switch to 'landscape' if the user wants horizontal), resolution='2K'. Prompt: dress the model from INPUT A in the outfit from INPUT B; take ONLY