Ditto
antgroup/ditto-talkingheadReviewed by AI
Checked by RepoPicks Radar on
What is Ditto?
Research code from Ant Group that animates a still portrait so it lip-syncs and moves along with an audio file you provide, saving the result as an MP4. You run it with one command after creating a conda environment, installing TensorRT and downloading about a dozen model files from Hugging Face. It was tested on Linux (CentOS 7.2) with an A100 card, and the ready-made fast models only cover NVIDIA Ampere cards or newer; older cards need a conversion step. There is also an official Colab notebook to try it in the browser.
Ditto is free and open source (Apache-2.0) and can replace HeyGen, Synthesia, and D-ID.
How easy is it to install Ditto?
D · terminal, several tools or a specific graphics card · Linux
On Linux you clone the project, build a conda environment, install TensorRT 8.6.1, fetch the model files with git lfs and then type a command with five paths; there is no window or page to click. Windows and macOS are not mentioned at all, and the tested setup was CentOS 7.2 with an NVIDIA A100.
Coming soon: install in 1 click with the RepoPicks app.
Ditto on GitHub
- Stars
- 932on Nov 10, 2026
- Downloads
- —all releases
- Growth
- +3stars in 6 days
Stars over time
Facts
- Runs on
- Linux
- You need to install
- Python, NVIDIA graphics card, FFmpeg
- To build from source
- Python, NVIDIA graphics card, FFmpeg
- Disk space
- 100 MB–300 MB · estimate
- Accounts or keys
- None
- Cost
- Free
- License
- Apache-2.0 (Permissive)
- Last change on GitHub
- Nov 12, 2025
Ditto risk signals (5)
- Attentioncan make a real person appear to speak
README.md describes "Controllable Realtime Talking Head Synthesis" driven by "--audio_path" and "--source_path" (any image). The same command that animates your own portrait will animate a photo of someone else with any audio, and the repository shows no consent check, watermark or usage restriction beyond the Apache-2.0 licence.
AI review
- Attentionno updates for a year
Last push on 2025-11-12, 360 days without commits, and no GitHub releases at all. The newest README entry is the training code published on 2025.11.12.
AI review
- Infoready models only for newer nvidia cards
README.md: "We have provided the tensorRT model with hardware-compatibility-level=Ampere_Plus (checkpoints/ditto_trt_Ampere_Plus/). If your GPU does not support it, please execute the cvt_onnx_to_trt.py script to convert from the general onnx model". On older cards you have to build the engines yourself.
AI review
- Infomodel files hosted elsewhere
README.md Download Checkpoints: "git clone https://huggingface.co/digital-avatar/ditto-talkinghead checkpoints". The model files, including a compiled plugin (libgrid_sample_3d_plugin.so), come from Hugging Face and not from this repository; the deterministic disk estimate (100 to 300 MB) covers only the code.
AI review
- Infopinned old dependencies
environment.yaml pins python=3.10.16, pytorch=2.5.1 with cuda 12.1, tensorrt==8.6.1 and numpy==2.0.1, all from late 2024, which can clash with a current driver or a newer CUDA install.
AI review
Evidence is shown in its original language.
What Radar could not check about Ditto
- Whether it installs and runs on anything other than the tested CentOS plus A100 setup, and whether it works on Windows at all.
- Real disk use after downloading the model files from Hugging Face.
- How fast it is on a consumer card, and whether the "realtime" claim holds outside an A100.
- How good the result looks with your own photo and audio.
- Whether the official Colab notebook still runs.