Local AI avatar videos as a HeyGen alternative: open models, licences and the workstation

Presenter videos from a photo and a script, on your own machine

A company can make HeyGen-style presenter videos on its own workstation with open models: InfiniteTalk or LongCat-Video-Avatar-1.5 animate a photo of a person so that lips, face and head follow a voice recording, and an open text-to-speech model reads the script aloud. By our arithmetic from the model files, one NVIDIA RTX PRO 6000 Blackwell Workstation Edition or Max-Q with 96 GB holds the whole chain in 16-bit precision for 720p, and an RTX PRO 5000 Blackwell with 48 or 72 GB holds it in compressed form. The photo, the voice, the script and the finished clips stay on the machine.

This guide is for the people who make these videos: HR and training teams with onboarding and safety courses, marketing teams with product explainers, and owners who present their company themselves. It covers the steps, the models a company may use under their licences, the limits to plan for, and which workstation fits which volume. The hardware sits in one section with a table near the end.

An avatar video, or talking-head video, shows a person speaking to the camera; here that person is generated from a still photo. Lip sync means the mouth moves in time with the words. A model is the trained AI itself, a set of files of several GB that the software loads onto the graphics card (GPU), and its licence says what a company may do with it.

Cloud services do the same in the browser. HeyGen's website, read on 30 October 2026, names "175+ languages and dialects" and video translation with voice cloning and lip sync. Synthesia's website states "140+ languages", personal avatars from "a single photo or a short video", ISO 27001 certification and "EU data residency". A local setup does without that ready catalogue of voices and languages. In return, no photo, voice or script leaves the company network, the number of videos is limited only by the time the GPU needs, and the models stay at the version you checked.

From script to finished clip in four steps

1. Script and voice. A text-to-speech model turns the script into an audio track. Chatterbox Multilingual from Resemble AI is under the MIT License, speaks 23 languages including German, English, French, Italian, Polish, Spanish and Swedish, and imitates a voice from a short reference recording; its card says every audio file it generates carries Resemble AI's Perth watermark. NVIDIA's Magpie TTS Multilingual is "ready for commercial use" under the NVIDIA Open Model License, covers 12 languages including German, and comes with preset voices; NVIDIA removed voice cloning from this release. A speaker can also record the script.
2. Face and lip sync. The avatar model takes the photo and the audio track and generates the video. InfiniteTalk, LongCat-Video-Avatar-1.5 and Wan2.2-S2V-14B are described below.
3. Upscaling. The avatar models generate 480p or 720p. SeedVR2 from ByteDance Seed, under Apache 2.0 and natively supported in ComfyUI, raises the picture to full HD; our guide to AI video upscaling compares the options.
4. Editing. Titles, slides and music go in with the video editor the team already uses.

Voice and face come from two different models with two licences, and both must allow commercial use. Our guide to on-premise voice agents covers the speech models in more depth.

Open avatar models and their licences, October 2026

GitHub repositories and Hugging Face model cards of MeiGen-AI, meituan-longcat and Wan-AI, the Wav2Lip repository and the LICENSE file of Tencent-Hunyuan/HunyuanVideo-Avatar, read 30 October 2026.

InfiniteTalk from MeiGen-AI was released with weights and code on 19 August 2025. It is built on the Wan2.1 image-to-video model with 14 billion parameters and uses chinese-wav2vec2-base as its audio encoder, a model whose card carries the MIT tag. The repository states: "The models in this repository are licensed under the Apache 2.0 License", and the authors claim no rights over the content you generate. It works from a single photo (image-to-video) or dubs an existing video (video-to-video), which the repository says "Enables unlimited length generation". For image-to-video it reports good results from a single image for up to one minute, and its default maximum is 1,000 frames, which it equates with 40 seconds. It warns that the FusionX LoRA, a speed-up add-on, makes colour shift over one minute worse and weakens the likeness.

LongCat-Video-Avatar-1.5 from Meituan is built on the LongCat-Video model with 13.6 billion parameters. Its card states that "The model weights are released under the MIT License", and that the licence grants no rights to Meituan trademarks or patents. Version 1.5 replaces the wav2vec2 audio encoder with Whisper-Large, generates 480p or 720p in 8 steps after distillation, and handles scenes with several speakers. The card's examples run on two GPUs with context parallelism, and the option --use_int8 loads a compressed model "for reduced VRAM usage".

Wan2.2-S2V-14B, released by the Wan team on 26 August 2025, turns a reference image and an audio track into video, optionally following a pose video, and the clip length follows the audio. Its card states Apache 2.0 and says its single-GPU command "can run on a GPU with at least 80GB VRAM".

Two models are out for a company in the EU: the HunyuanVideo-Avatar licence excludes the European Union, and the Wav2Lip README limits use to "research/academic/personal purposes". We install neither. The same territory clause covers HunyuanVideo 1.5, as our guide to video generation GPU requirements shows.

The limits to plan for in language, length and picture

German and other languages. InfiniteTalk's audio encoder was pretrained on 10,000 hours of the WenetSpeech corpus of Chinese speech, and LongCat-Video-Avatar-1.5 reports its human evaluation in Chinese and English only. We found no published evaluation of either model in German, French or another European language. Generate a test clip in each language you need and have a native speaker review the lip movements before you plan a course series on it.

Length. For a single photo, InfiniteTalk's guidance is good results for up to about a minute. A ten-minute course is therefore made as one clip per slide or section, each under a minute, joined in the editor, which also gives the viewer natural breaks.

Resolution. 480p suits drafts and small embedded players; 720p plus an upscaler gives a full-HD course video. Resolution costs memory and time, because the model processes every frame as a long sequence of image patches. Our Wan 2.2 server article shows the arithmetic for Wan 2.2: a 720p clip takes about four times the compute of a 480p clip.

Speed. None of the publishers states how long a clip takes on an RTX PRO card, and we found no independent measurement. The time depends on resolution, clip length, the number of steps (8 for LongCat-Video-Avatar-1.5) and whether the weights are compressed. Have one clip of your usual length timed on the delivered workstation before you commit to a production schedule.

You can order the workstation alone, the setup of InfiniteTalk, LongCat-Video-Avatar-1.5, a voice model and ComfyUI alone, or both together: the workstation delivered with everything installed and configured; write to us and we suggest the best solution for your tasks and budget.

Consent, GDPR and the AI Act

An avatar of an employee uses that person's face and voice. The European Commission defines personal data as "any information that relates to an identified or identifiable living individual" and lists posting a photo of a person on a website as an example of processing. The cloud services build a consent step into their product: HeyGen states that "Every custom avatar requires on-camera identity verification from the person depicted", and Synthesia that "Every Personal Avatar needs a live, on-camera consent recording". A local workstation has no such gate, so the company sets the rule itself: a written agreement with the person shown, a stated purpose, and a way to withdraw that also retires the videos. For employees, the Commission notes that consent is not freely given where there is a clear imbalance, for example between employer and employee, so your data-protection officer confirms the legal basis; our guide on local AI and the GDPR explains what own hosting changes and what it does not.

The AI Act defines a deep fake as "AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful" (Article 3(60)). Under Article 50(4), deployers of a system that generates deep fakes "shall disclose that the content has been artificially generated or manipulated"; for evidently artistic, creative, satirical or fictional works, that duty is limited to disclosure that does not hamper the work. Article 50 applies from 2 August 2026, as our article on the EU AI Act explains. Chatterbox marks its audio with a watermark, but the disclosure to viewers is the company's task, at the latest when they first see the video (Article 50(5)). A company that sets up such a system for its own use can also count as its provider, and providers mark generated audio and video in a machine-readable format (Article 50(2)); the pages of InfiniteTalk and LongCat-Video-Avatar-1.5 describe no such mark for the video. From 2 December 2026 Article 5(1)(ba), added by Regulation (EU) 2026/1744, also prohibits AI systems that generate intimate or sexually explicit images, video or audio of an identifiable person without that person's explicit consent. Your legal counsel confirms what your videos need.

Memory for the avatar models and the workstation tiers

No publisher of these models states a memory figure, except Wan's 80 GB for its own S2V command. We therefore add up the files the software loads, as listed on Hugging Face. Working memory for the frames comes on top and grows with resolution and clip length.

Hugging Face file lists of Comfy-Org (Wan 2.1 and Wan 2.2 repackaged), WanVideo_comfy (InfiniteTalk module, 16-bit only) and meituan-longcat/LongCat-Video-Avatar-1.5, read 30 October 2026; chain totals our arithmetic, including the 0.25 GB Wan 2.1 VAE; image and audio encoders not counted.

our sizing from the file table above and the model cards; card memory and power from NVIDIA's datasheets.

Entry. The RTX PRO 5000 Blackwell has 48 or 72 GB at 1,344 GB/s and draws 300 W. The FP8 chain of about 28.5 GB leaves room for 480p frames on the 48 GB card, and the 72 GB card also holds the 16-bit LongCat model. A 32 GB RTX PRO 4500 runs InfiniteTalk only if the software moves the text encoder off the card after reading the prompt, which leaves about 21.8 GB of weights by our arithmetic; nobody has published whether the frames then fit, so we treat it as a card for trials. The 24 GB RTX PRO 4000 is below the FP8 chain. These margins use our usable-memory rule (90 per cent of the memory the driver reports, less 3 GiB): about 40.0 GiB on a 48 GB card, 61.5 GiB on 72 GB and 25.7 GiB on 32 GB.

Mid. One RTX PRO 6000 Blackwell holds the 16-bit chains, meets Wan's 80 GB for S2V and keeps the voice model on the same card. The Workstation Edition draws 600 W, the Max-Q 300 W with a blower cooler. A second card lets two clips render at once or matches LongCat's two-GPU example. Two Workstation Edition cards are the realistic maximum for that edition, at about 1.8 kW at the wall by the estimate in our article on two RTX PRO 6000 Workstation Edition; two Max-Q come to about 1.1 kW, as our Max-Q power guide estimates. Three or four Max-Q belong only in a chassis built for them with strong directed airflow: Puget Systems (July 2025) found that directly adjacent Max-Q cards have little room to draw air, and Exxact (updated March 2026) reports that four adjacent cards throttled without its custom airflow.

Large. A server with four RTX PRO 6000 Server Edition cards gives a department one clip per card. InfiniteTalk documents multi-GPU runs with FSDP and Ulysses, eight GPUs in its example, but publishes no scaling figures for cards linked only over PCIe. By our estimate such a server draws about 3.4 kW and needs a 32 A or three-phase feed in a server room, as our page on the four-card Server Edition server explains.

ComfyUI as the interface, and what stays on the workstation

When we set up the software, the workstation gets ComfyUI, under GPL-3.0 and actively developed as of October 2026, and the ComfyUI-WanVideoWrapper under Apache 2.0, which InfiniteTalk's repository credits for its ComfyUI support. ComfyUI's App Mode, which NVIDIA's blog of 10 March 2026 calls App View, presents a workflow as a simple form: the trainer picks the photo and the voice, pastes the script and starts the job, without touching the node graph. Wan2GP, another interface for these models, is not part of the software we set up: its WanGP Community License 2.0 requires a separate licence for "embedding the Software in a paid product or paid service".

Photos, recordings, scripts, generated audio and video, and the job history stay on the workstation. The internet is needed only to download models and software updates; generating a clip runs offline. ComfyUI states a weekly release cycle, so decide within your team who updates the installation and re-run a reference clip after each update.

We supply avatar workstations from one RTX PRO 5000 to two RTX PRO 6000 Workstation Edition or Max-Q and servers for larger teams; send us the number of videos per month, their length and languages, and we propose the configuration and the models for them.

FAQ