kunal12203/higgsfree

higgsfree is an open-source pipeline for generating a talking-head video from a short consent video and a text script. You give it a video of a real person and the words you want them to say, and it produces a photorealistic video where that person appears to speak the script, with their face preserved, their voice cloned from the original recording, and their lips synced to the generated speech. The pipeline runs nine stages in sequence. It extracts the best face frame from the consent video, generates a portrait image using AI models that preserve the person's facial identity, extracts a voice profile from the audio, synthesizes speech in that cloned voice, and then applies lip-sync animation so the mouth movements match the generated speech. A final face restoration step polishes the result, and the talking head is composited onto a background scene before the audio and video are combined into the final file. Three pipeline variants are available. One produces a full seated studio portrait with a scene background. One outputs just the talking head with minimal setup. A third generates video from a text description alone, without any source person. The scene options for the avatar variants include studio, cafe, outdoor, and desk backgrounds. The project is designed for contributors: each model runs in its own isolated environment so dependencies do not conflict, every stage caches its output so a re-run resumes from where it left off, and there are fallback options at each step in case the primary model fails. Quality is scored automatically on every pull request using face identity similarity and lip-sync confidence metrics. Running it requires an NVIDIA GPU with 16 to 20 gigabytes of video memory, CUDA, FFmpeg, and Python 3.10 or newer. Docker is also supported. The license is MIT.