ReShoot 4D: How to Change Camera Angles in Existing Video With AI
ReShoot 4D rebuilds 3D scenes from footage so you can reshoot with new camera moves. Here's how its 4-step ComfyUI pipeline works.

What is ReShoot 4D?
ReShoot 4D is a free, open source tool that takes a video you already shot and lets you reshoot it from a different camera angle, orbiting around a subject, pushing in for a close-up, craning upward, or pulling off a bullet-time style freeze move. Instead of generating a new scene from a text prompt, it reconstructs the 3D geometry of your existing footage and then renders a virtual camera moving through that reconstructed space. It was built by a single developer, who has said the tool was created with help from Claude Code and Codex, and it runs through ComfyUI on Nvidia GPUs.
TL;DR
- ReShoot 4D reconstructs a 3D scene from 2D video, working out where the original camera sat and how the subject moved through time, before letting you place a new virtual camera inside that space.
- The pipeline runs in four steps: 3D scene capture, virtual camera placement, re-rendering from the new angle, and a final cleanup pass using an open video model called LTX to smooth out artifacts.
- It requires an Nvidia GPU and familiarity with ComfyUI, which means it’s not a drag-and-drop consumer app, it’s a node-based workflow aimed at people comfortable with that ecosystem.
- The tool doesn’t invent what it can’t see, so anything outside the original camera’s frame can shift or change inconsistently when you swing the virtual camera to a new position.
- Fast motion can smear faces for a few frames, a known limitation the developer has flagged, so the current results work best on relatively controlled, slower footage.
- A public demo lets you test the camera editor on preloaded footage, but direct upload of your own video isn’t available yet in the version shown.
- It sits alongside a wave of other camera-and-scene tools, including a similar reshoot concept previously demoed by ByteDance on an old movie scene, though that one wasn’t downloadable.
- Self-hosted
- SOC 2
- HIPAA
- Audit log
How does ReShoot 4D actually work?
The tool breaks the process into four distinct stages.
First, it builds a four-dimensional capture of the scene, meaning a 3D reconstruction that also accounts for how things move across time. This step figures out where the original camera was positioned throughout the shot and tracks how the subject moved relative to it.
Second, you define a virtual camera. You can manually set its path, lens, and focus, or choose from presets like an orbit around the subject or a crane move that rises vertically. This is the creative control layer, the part where you decide what the “reshoot” actually looks like.
Third, the system renders the scene from that new virtual camera position. Because it’s working from a reconstructed 3D approximation rather than a fully complete world, this raw render often comes out with visible fuzziness and artifacts, especially around edges and backgrounds.
Fourth, a cleanup pass using LTX, an open video model, smooths over those rough edges so the final output looks closer to real footage rather than an obvious synthetic render.
What are its hardware and software requirements?
ReShoot 4D runs through ComfyUI, the node-based interface widely used for running diffusion and video models locally. That means setup isn’t plug-and-play. Users need an Nvidia graphics card, and ComfyUI itself has a learning curve that’s steeper than most consumer apps, involving workflow files, model downloads, and node graphs rather than a simple upload button.
Specific VRAM requirements and exact GPU models weren’t detailed, so anyone evaluating whether their setup can handle it should check the project’s own documentation before assuming compatibility.
What are the current limitations?
A few constraints matter if you’re thinking about using this for real editing work.
The tool can’t generate content it never saw. If part of the original scene is out of frame, say the far side of a room or what’s outside a window, the reconstruction has to guess or approximate that area when the new camera angle reveals it. In practice, this means background details can shift inconsistently between different camera positions in the same reshoot, since the model is filling gaps rather than recalling something it actually captured.
Fast motion is another weak point. The developer has noted that quick movement can cause faces to smear across several frames, which limits how well this works on action-heavy or handheld footage versus something more static, like a dialogue scene or a slow push-in shot.
At the time this was covered, direct upload of personal footage wasn’t available in the public demo. Users could manipulate camera paths on preloaded example clips to see how the orbit, crane, and other presets behave, but couldn’t yet run their own video through the full pipeline in that interface.
Is ReShoot 4D worth trying right now?
For anyone doing previsualization, shot planning, or experimenting with camera language in post, it’s a genuinely useful free tool, provided you have the right GPU and some patience for ComfyUI. The fact that it’s open source and runs locally (rather than being locked behind a paid cloud service) makes it accessible to people who want to tinker without a subscription.
That said, it’s not a finished production tool. The artifacts introduced during the raw re-render step, the inconsistencies in off-camera content, and the face-smearing issue on fast motion all mean this is better suited to experimentation and lower-stakes projects than final deliverables for clients or paid work. It’s also a one-developer project, which affects how quickly bugs get fixed or features get added compared to something backed by a larger team or company.
How does this compare to other camera-and-scene AI tools?
ReShoot 4D isn’t the only project tackling this problem. A similar concept, reworking camera angles in existing footage, was demoed by ByteDance on an old movie scene, though that version wasn’t available for public download at the time. ReShoot 4D’s distinction is that it’s actually downloadable and open, even if it requires technical setup.
It’s also conceptually different from AI world models like Tencent’s Hunyuan World Model, which generates entirely new, explorable 3D environments from scratch and lets a user navigate a camera through a world that never existed as footage. ReShoot 4D works in the opposite direction: it starts from real video and tries to preserve what actually happened, just viewed from a new angle. Both rely on reconstructing or generating 3D structure to support camera movement, but one is about creating new worlds and the other is about re-examining old ones.
Frequently Asked Questions
What video formats or lengths does ReShoot 4D support?
Specific supported formats and maximum clip lengths weren’t detailed in available coverage. The demo shown involved short preloaded clips rather than long-form footage.
Do I need a powerful GPU to run ReShoot 4D?
Yes, it requires an Nvidia GPU to run through ComfyUI, though exact VRAM or model requirements weren’t specified. Anyone planning to use it should check the project’s own setup documentation before investing time in installation.
Can ReShoot 4D create camera angles for parts of the scene that were never filmed?
Not reliably. The tool reconstructs 3D structure from what’s visible in the original footage, so anything that was off-camera has to be approximated rather than accurately recreated, which can cause inconsistencies across different virtual camera angles.
Is ReShoot 4D free to use?
Yes, it’s described as free and open source, built by a single developer rather than a company, and distributed for anyone with the right hardware and ComfyUI setup to run it.
How is this different from just filming multiple camera angles on set?
It’s meant for situations where you only have one take or one angle already recorded and can’t go back and reshoot. Instead of physically moving a camera during production, you reconstruct the scene afterward and generate a new synthetic camera path, with the tradeoff being visible artifacts and limitations compared to footage actually filmed from that angle.