A video character replacement has several inputs worth inspecting before the final generation: the source performance, the character reference, and the guidance that tells the model which subject to change.

I built Henge, an open-source workspace that puts those stages into a visible prepare → review → generate workflow.

Henge source and replacement comparison

Source code and installation · Watch the 20-second demo

The demo compares a source clip with the actual result of one replacement run, using original AI-generated actors. The promotional edit omits processing time; the repository also includes the uncut result.

Prepare, inspect, then generate

Start with a short source video and a character reference. Henge prepares motion guidance, lets you inspect it, and ties approval to that version of the inputs.

There are three guidance methods: Depth + SAM, Fast Face Mesh, and a combination of Depth + SAM + Face Mesh. The face-mesh methods require an optional worker and are aimed at a single centered speaker.

During review, check the full clip for hair, hands, clothing, mask coverage, and spill into the background. Changing the prepared input invalidates the earlier approval. Once the inputs are ready, submit the final generation, preview the output, and export with the source video's original audio restored locally.

The application runs on your machine, while AI processing uses paid cloud services. You need your own DreamAPI and Replicate keys and available credit. Guidance preparation can incur charges too; review is a checkpoint before final generation, not a promise of a free preparation stage.

Recovery is part of the workflow

A long-running generation can finish even after the browser closes or the local connection breaks. Henge saves provider task IDs so reopening a task can reconnect to work that has already been submitted.

An uncertain submission blocks automatic resubmission. A timeout without a task ID needs a check of provider history and billing before deliberately starting another paid job. This makes the state of a task visible instead of treating every interruption as a reason to generate again.

Keys entered in the application use macOS Keychain or Windows Credential Manager. Source derivatives and references are processed by the selected cloud providers; this is a local interface, not an offline generation pipeline.

What is open source

Henge is the workflow and interface. It does not include an open-source video generation model.

DreamAPI provides the Seedance 2.5 generation integration. Replicate supplies preparation services including depth, vocal separation, and the current SAM adapter. The repository documents those dependencies and their terms.

Henge is derived from Sirio Berati's Genjustsu Open Source Workflow, with the upstream MIT license retained. Upstream example videos and branding are not distributed with Henge.

Try a short, single-subject clip

On macOS, you need Python 3.12+ and Rubber Band:

git clone https://github.com/Hy-1990/Henge.git cd Henge brew install python@3.12 rubberband PYTHON=python3.12 bash setup.sh bash start.command 

Open http://127.0.0.1:8771 and configure the two provider keys. Windows setup instructions are also included in the repository.

Start with a clearly visible single subject in a 4–30 second clip. Automatic identity matching across cuts and pixel-perfect preservation of every non-target person are not guaranteed. A live macOS Depth + SAM / image-reference / 720p workflow has been verified; other modes and resolutions still need their own quality checks, and a paid Windows generation run has not yet been verified.

The repository includes the source clip, character reference, uncut output, setup instructions, and current limitations.