Library
All captures and outputs
No performances match the current filter.
Preparing FBX download…
Versions
Source capture
Original recording
Apply to character
Hybrid-WAN auto-recommends the newest actor Wan LoRA when a face-ref video is present. Base model remains available as an explicit fallback.
Auto uses model default prompt (recommended).
0 = quality, 1 = fidelity. Default 0.5.
Stage 8 will be skipped.
LTX-2.3 background plate. Producer-only — not composited yet.
Hybrid-WAN requires you to click the exact character reference to use.
Generated
Voice Dubbed
Fish Speech clone
No character generations yet.
Start camera to preview
Scan with phone…
Recorded or selected take appears here
Scan with your phone:
If this opens in DuckDuckGo, use Safari for LAN links.
Waiting for connection...
Connect your phone camera
Drop video here
MP4, MOV, WebM · Max 200MB
Send from phone
Waiting for video...
Loading QR...
Uploading...
Camera 1
Camera 2
Use a printed ChArUco board and record until all target zones are covered in both cameras.
Calibration target: per take. Use the printable 10×7 ChArUco board (9×6 inner corners, 25mm squares, 18mm markers). Minimum quality: , per camera.
The board must be visible in both cameras at the same time, but it does not need to be in the same zone in both views. Zone overlays turn green after analysis. If preview looks mirrored on your device, continue anyway; calibration uses board geometry, not image orientation.
- Keep both cameras fixed and angled toward the same play area (roughly 20-45 degrees apart).
- Hold the board so both cameras see all corners in each sampled moment.
- For each camera, cover all four zones over time (zones can differ between cameras per moment).
- Tilt/rotate slightly (about 15-30 degrees), and vary distance (near and far).
- Use a rigid board backing (cardboard/foam core). Curled paper reduces calibration quality.
- Avoid motion blur; pause briefly (0.5-1s) in each zone.
The timer is guidance only. You can record longer than target and add multiple takes; results are aggregated.
Sweep the calibration board through each highlighted zone and pause briefly when stable. Recommended minimum: . No auto-stop.
Output
Select a file to preview output.
Upscale
Next Step
This capture is loaded into Generate for the selected character.
Generate Pipeline
Click the exact ref for Hybrid-WAN.
One click flow: extracts poses first, then runs generation.
Official mode skips DWPose extraction and runs the upstream Wan preprocess during generation.
Drag a saved clip onto the timeline to append it as a segment.
Record or upload a video first
Click to view intermediate stage outputs for A/B comparison.
Hover to preview. Character over environment.
No library videos yet
Hybrid-WAN recommends the newest actor Wan LoRA for identity lock when a face-ref video is attached. You can still switch back to the base model here.
Locked baseline using upstream YOLOv10m + ViTPose preprocessing, upstream face conditioning, retarget on, 30fps preprocess, no Wan LoRA, and no custom prompt overrides.
Auto mode sends no custom prompt and lets Wan2.2 use its default conditioning prompt.
Negative prompt has strongest effect when Prompt Strength is above 1.0.
Hybrid-WAN is locked to an explicitly chosen reference. Click the exact image you want.
Set a primary ref on the Actor page to enable auto-select
Share Video to Project
No projects available. Ask an admin to create one.