EchoMimic V3 Sampler

This is the node that does the actual "talking." The EchoMimic V3 Loader hands it a pipe; you hand it a reference photo and some audio; it hands back a stack of video frames of that face talking, in sync, with head motion and expression that look close to human. That last part is the whole reason to use EchoMimic V3 over older talking-head models - the head actually moves naturally instead of hovering like a haunted portrait.

How it works

EchoMimic V3 is audio-driven diffusion built on a Wan2.1-Fun backbone. The sampler runs your audio through the wav2vec2 audio encoder to get speech features, then conditions the diffusion model on those features plus the reference image, frame after frame. The two guidance knobs - cfg for how hard the visual diffusion listens to the prompt, audio_cfg for how hard it listens to the audio - are where you balance "mouth moves" against "everything else stays stable."

Inputs that matter

• ref_image - feed it a clean, front-facing photo, roughly square. The README says 1:1 (like 768×768); heavily cropped or angled shots make the model invent features.
• ref_audio - the drive track. Any AUDIO node output works.
• prompt / negative_prompt - the negative prompt ships preloaded with a serious eye-correction block ("rolling eyes, bad eyes, 眼神飘忽..."). Leave it alone; if your output has the model staring into the middle distance, that prompt is the lever, not your positive prompt.
• steps / cfg / audio_cfg - defaults of 25 / 4.0 / 2.9 are sane starting points. If the mouth lags the audio, nudge audio_cfg up.
• fps - default 25. Set it to what your downstream video combine expects.

The optional group has the performance controls: enable_teacache (off by default in the schema - flip it on for roughly 1.5× speed at near-zero quality cost; the README claims it's on by default, but the shipped schema disagrees), teacache_threshold (0.1 is fine), context_length (the temporal window, default 49 - lower it if you're VRAM-starved), and keep_model_loaded (on by default: keeps the model resident on the GPU so consecutive generations don't reload, at the cost of VRAM).

Output and what it plugs into

One output: images (IMAGE tensor). Send it to AIIA Video Combine (same pack) to mux it into a real video file, or straight to a preview/save node. Note this is the in-memory path - for very long audio the frames pile up in RAM, which is exactly why the pack also has disk-streaming options elsewhere.

Troubleshooting

• Eyes drifting or rolling - your negative prompt got stripped or changed. Restore the default.
• Mouth out of sync on long clips - EchoMimic V3 is at its best on short-to-medium audio; for marathon segments the pack's own docs recommend cutting the audio and merging, or moving to the disk-based flow.
• OOM - drop context_length, turn off keep_model_loaded, and keep the reference image modest. This is a 1.3B diffusion model; 24GB of VRAM is the comfortable lane.