EXPLORE BFS V3 / LTX-2.3

A new face.
The same
performance.

One portrait. One guide video. Explore a different identity while keeping the motion, scene, and original audio.

Prepare your inputs ↗Watch model examples ↓

REFERENCE → VIDEOOPEN MODELV3 WORKFLOW

FROM THE MODEL CREATOREXAMPLE 06

Published model example · includes the creator’s comparison layout

01 / YOUR WORKSPACE

Set the scene.

GPU access pending

Input preparation is ready. Video generation is waiting for a GPU.

You can preview your V3 guide and download a setup file here. This free page does not run the model. Your selected files stay in your browser.

GPU grant request ↗

Reference face

IMAGE

＋Add a portraitA clear, well-lit face works bestJPG, PNG or WebP · up to 20 MB

The identity you want to use

Guide video

VIDEO

▷Add a performanceA short, single-person close-upMP4, WebM or MOV · up to 150 MB

The motion and scene you want to keep

Clip length Canvas resolution Seed

Describe the reference & scene ＋

Reference identityPerformance and scene

V3 uses a structured FACE / ACTION prompt. Use the portrait for identity traits and the video for the performance.

◈ Prepared on your device. No uploads, account keys, or GPU time.

Guide preview

LOCAL

▧

Your next performance
starts here.

Add a portrait and a guide video
to preview the V3 reference layout.

This is the conditioning guide, not a generated face swap. The reference portrait stays visible in a green side panel throughout the video.

Generate video — available after GPU activation

02 / SEE WHAT’S POSSIBLE

From the model creator.

All model examples ↗

Consistent identity

V3 / EXAMPLE 07

A reference that stays

V3 / EXAMPLE 08

These are prerecorded outputs published by Alissonerdx, not generations from this Space.

THE WORKFLOW

One identity. Every frame.

BFS V3 places your reference in a persistent side panel, giving the model an identity signal throughout the performance. The model learns to return the scene without the reference strip.

A BETTER FIRST TAKE

Keep it clear and close.

Start with one visible face, steady lighting, and a short clip without cuts. Fast motion, occlusion, and hair differences can cause drift or distortions. Use media and likenesses you have permission to use.

READY FOR WHAT’S NEXT

A GPU-ready backend.

The Space includes a Gradio app with BFS V3 rank-64, LTX-2.3, original-audio preservation, and downloadable MP4 results. Inference has not been tested while GPU hosting is unavailable.

Explore the source ↗