Ditto

Ditto (ditto-talkinghead) animates a character's picture from the voice it speaks with, so the picture talks, blinks and moves instead of sitting still. It works on photos and on drawn characters, and keeps a cutout's transparent background.

Marked experimental, and only administrators can add it.

Requirements

A GPU with about 5.5 GB of free VRAM (the models are a ~2.2 GB download).
A picture with a face Ditto can find. Pictures without one stay still.

Setup

Add the service

Manage Modules → Add Modules → Ditto → Install. Voxta installs the Python runtime on first use.

Download the models

In the service settings, keep the suggested Model and press Download to fetch it from HuggingFace (into the Ditto models folder by default).

Enable it for your chats

In the services configuration, turn on Portrait Animation and pick Ditto. See Portrait Animation.

Presets

Preset	Between lines
Ditto (Default)	About ten seconds of quiet motion are rendered once per picture and replayed. No further GPU cost.
Ditto (Live Idle)	The face keeps rendering at half rate while the chat is open. Livelier, but keeps a good part of the GPU busy.

Settings

Setting	Default	What it does
Maximum Picture Size	1920	Longer pictures are shrunk before the face is found. The face is always animated at 512 px and laid over the original, so this only affects sharpness around the face, not speed.
Prepared Pictures	16	Pictures kept prepared in memory (~10 MB each). Each expression is its own picture: cover characters × expressions in a scene. Preparing one takes a few seconds.
Motion At Rest	35%	Head motion kept while nothing is said. 0 holds the picture's pose; 100 moves as much as while speaking.
Motion Smoothing	3	Frames the head and mouth motion is smoothed over. Higher is steadier; 1 is raw model output.

Limits

One character speaks at a time. A second request waits and the picture stays still meanwhile.
Runs faster than real time on a recent GPU.

Ditto

On this page