Introduction

What if you could go back and move the camera after the shoot? That's the pitch behind Reshoot 4D, a new open-source ComfyUI node pack from developer itsmetamike, posted to X at 07:55 UTC on October 30 as an entry in Comfy's #OpenSourceHiggsfield challenge. Feed it a video of one person, and it builds a 4D capture of them in their setting, opens a camera editor right inside a ComfyUI node, and reshoots the clip along the new path you draw. LTX 2.3 renders every final frame.

It runs locally on open weights, with no API keys, no cloud services, and no partner nodes. The code is AGPL-3.0 and lives at github.com/itsmetamike/reshoot-4d, with a ready-to-load Reshoot 4D.json workflow. The author tested it on Ubuntu 24.04 with a single RTX 3090 (24 GB), where one shot takes about 30 minutes from camera move to final video.

The Four Nodes

The whole pipeline is four nodes, wired in a line:

  • Capture turns your video into a 4D capture of the person in their setting and solves the real source camera. Under the hood it uses Ant Research's 4DAnyone, GVHMR body tracking, and Depth Anything 3. The 4DGS mode also trains a Gaussian splat for the editor preview.
  • Virtual Camera opens the 4D camera editor on the node itself. You keyframe a new camera over the moving capture, and the path saves with the workflow.
  • Retake reshoots the whole source video along your path with ReDirector, built on Wan 2.1 Fun 1.3B Control-Camera. It samples 81 frames. Start at 480p, 50 steps, CFG 5.
  • Photoreal Pass hands the retake to LTX 2.3 (22B distilled) with Cseti's CrossView-Warp IC-LoRA. MoGe-2 estimates the real video's depth, the real footage gets warped along your path, and LTX finishes the frames at up to 1792×1024 (or 1024×1792 portrait), with the LTX detailer on by default.

That last stage became the default only today. The commit at 07:01 UTC on October 30 switched to "CrossView + detailer as the default Photoreal Pass" and re-rendered every result.

The Camera Editor

The Virtual Camera panel is the fun part, and it feels a lot like a mini Blender:

  • Camera keys for position, aim, and roll, joined by smooth orbits or splines
  • Camera view vs. editor view: press C to step outside the shot and see the capture, the path, and every key
  • A Blender-style view gizmo for top, front, and side views
  • A shot timeline with J / K / L shuttle, frame stepping, and multi-key retiming
  • Presets for arc, push, crane, orbit, and bullet-time moves
  • Portable paths: moves export relative to the source camera, so one move can replay on another capture

There's also a browser demo of the editor on the project page if you want to poke at it before installing anything.

Install

Clone it into custom_nodes, along with the pinned CrossView-Warp node and the ReDirector installer:

cd ComfyUI/custom_nodes
git clone https://github.com/itsmetamike/reshoot-4d.git
git clone https://github.com/cseti007/ComfyUI-CrossViewWarp.git && git -C ComfyUI-CrossViewWarp checkout 3266a83a84eaa6833a9ebc38ac7cdd8789e44f5b
bash reshoot-4d/scripts/install_redirector.sh

Be ready for a real setup session. The install guide uses several separate Python environments (Capture, the splat trainer, ReDirector, and ComfyUI itself), CUDA 12.8 tooling, and a long list of pinned model files. Restart ComfyUI afterwards and you'll find the nodes under Reshoot 4D.

Limits and Licenses

The README is refreshingly honest about what doesn't work yet:

  • No fixed world. It doesn't build a 3D set and composite the person onto it. Anything the source camera never saw gets regenerated, so off-camera objects can change between angles.
  • Fast motion smears. Quick spins and fast arms can melt a face for a few frames.
  • Start small. The workflow guide suggests modest arcs or pushes first. Big moves to the unseen side need checking.

Licensing is mixed, so read the models and licences page before you use it for paid work. The node code is AGPL-3.0 and several models are Apache-2.0 or MIT. But the current Capture path includes GVHMR (education, research, and non-profit use only), SMPL-X (non-commercial research terms), and a CC BY-NC-SA 4.0 view-generation LoRA. LTX 2.3 comes under the LTX-2 Community License.

If you read our earlier piece on 4DAnyone, this is what it looks like when someone wires that capture step into a full director's chair.

Original Source

Conclusion

Reshoot 4D isn't a one-click toy. It's a 24 GB, half-hour-per-shot pipeline with a stack of non-commercial pieces. But it does something most video tools still can't: it takes real footage and lets you choose a new camera move after the fact, all inside ComfyUI on your own GPU. Start with a short, slow clip and a gentle orbit, and see how far you can push the angle.

—Aurelia ♡