Reshoot 4D

Reshoot any video in 4D, inside ComfyUI.

  • Runs locally
  • Open weights
  • Every step a ComfyUI node
  • Final frames by LTX 2.3

How it works

Why 4D

You direct inside a 4D capture of the real scene: real scale, the real source camera path, the person's real motion and timing. The retake is driven by that geometry, and the pixels come from reshooting the real footage, so the person, place and motion stay theirs. LTX 2.3 renders every final frame. The capture is also a reusable asset, ready for video models that take 3D conditioning.

Inside ComfyUI

The workflow and the 4D camera editor, from video to final shot. Try the editor in your browser →

Results

Each example: the source video, then three new camera moves. The top view on each reshoot shows the path; the readouts show how far it moves.

Limitations

Example 4: source and two reshoots; the view through the window differs
No fixed world. There is no 3D world built and composited behind the person. Each reshoot regenerates what the source camera never saw, so off-camera objects can differ between angles. In Example 4 the view through the window changes.
Example 6: source and reshoot mid-spin; her face smears in the reshoot
Fast motion smears. Quick spins and fast arms can melt a face for a few frames. In Example 6 her face smears mid-spin.

Install

  1. Set up Capture and its models on a CUDA machine.
  2. Add Reshoot 4D to ComfyUI/custom_nodes/.
  3. Install ReDirector and LTX 2.3, then restart ComfyUI.

Copyright (C) 2026 itsmetamike. AGPL-3.0. Models keep their own licences; see the model guide. Built with assistance from Claude Code and Codex.