Recast — Video Character Replacement · Open Skill 公開スキル

Recast — Video Character Replacement

Recast turns an existing video into a new character-driven version. Upload a source video and a character image. Recast uses the video as the motion and performance reference while using your uploaded character as the visual identity reference. Your new character can follow the original video's movements, poses, actions, timing, camera behavior, object interactions, and environment — without manually recreating the scene. HOW IT WORKS 1. Upload a source video. 2. Upload the character you want to use. 3. Recast analyzes the motion and character reference. 4. Review the production plan. 5. Approve generation. 6. Recast generates the video and checks the result. GREAT FOR • Character replacement • AI character videos • Recasting existing motion • Social media videos • Shorts and vertical content • Original characters and mascots • Stylized or realistic characters • Reference-driven video creation SMART REFERENCE CONTROL Source Video = Motion Authority Character Image = Identity Authority Recast preserves the source video's performance while prioritizing your uploaded character for appearance. It can preserve: • Body movement • Action sequence • Timing • Poses • Camera behavior • Framing • Environment • Object interactions • Overall pacing while replacing the original performer with your character. IDENTITY-FIRST WORKFLOW Recast checks whether your character image can be used directly. When possible, it uses your original reference without unnecessary regeneration, helping reduce identity drift. SAFE GENERATION Recast plans before generating and uses controlled generation behavior. It does not automatically keep regenerating videos when something goes wrong. Upload a video. Upload your character. Recast the performance.

@metebey

TryOn

77 9 39

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/223892/2026-09-24/jfegny6b.png

# Recast — Video Character Swap

## When to use
- User uploads a SOURCE VIDEO + a NEW CHARACTER IMAGE and wants to replace the main performer with the uploaded character, preserving the source motion, action sequence, timing, staging, camera behavior, composition, object interactions, and scene.
- Simple requests like "Replace the person in this video with my character."
- V1 is focused on character swap only — NOT dance, ads, general editing, environment/product replacement, full remix, or choreography.

## Core promise
"Upload a video and a character. Recast the video with your character."

## Reference files (read on demand, don't read all at once)
- `references/workflow.md` — read first; input collection, source analysis, Production Summary, planning/generation separation.
- `references/character-reference.md` — strict identity authority, direct-use gate, cropped-reference preparation, strict replica rules.
- `references/ref2v-generation.md` — source video = motion authority, character image = appearance authority, ref2v pipeline, motion fidelity, object interactions, camera/environment preservation.
- `references/qc.md` — identity fidelity, source-performer leakage, continuity, motion, interaction, camera, environment, artifacts, PASS/FAIL definitions.
- `references/safety-and-generation-control.md` — Safe Mode, one approval = one call, no auto retry, failure behavior, anti-loop.

## Steps
1. Read `references/workflow.md` and follow it. Use write_todos to track progress.
2. Collect inputs: SOURCE VIDEO + NEW CHARACTER IMAGE (ask only for the missing required input).
3. Analyze the source video (recognize_video) → SOURCE MOTION MAP; analyze the character (recognize_image) → CHARACTER IDENTITY LOCK.
4. Determine CHARACTER REFERENCE STATUS: DIRECTLY USABLE or PREPARATION REQUIRED (character-reference.md).
5. Present a Production Summary and get planning approval before generating.
6. On approval, generate per ref2v-generation.md (model='ref2v'). QC once (qc.md). Report. STOP.

## Generation conventions
- generate_video with model='ref2v' (reference-to-video): pass the NEW CHARACTER IMAGE and the SOURCE VIDEO in input_urls (character = appearance authority, source video = motion authority), plus a prompt that enforces authority separation. At least 1 reference image is required.
- generate_image (model='i2i.hd.lite') ONLY for strict character preparation when the reference is not directly usable; never for the core recast.
- name/entity_type: generate_video → name = segment title, entity_type='final_videos'; generate_image → name = 'Character Preparation', entity_type='reference_frames'; create_document → name = 'Production Summary', entity_type='plans'.
- Video generation does not support NSFW/adult content on any channel — decline such requests.

## Notes
- STRICT authority separation: NEW CHARACTER IMAGE = appearance; SOURCE VIDEO = motion/performance/camera/staging. Never preserve the source performer's identity.
- Planning approval and generation approval are separate; never generate while the user reviews the plan.
- One approved generation = one call = one QC = report = STOP. QC failure is not permission to regenerate.
- Report failures honestly; never fabricate URLs.