Master settings guide · Updated 5 October 2026 · Benchmarks from 4 October

SadTalker best settings: choose by goal, then check the output

Start with the framing and head movement you need. Choose the base render size, then decide whether face restoration helps. These presets bring together our four ComfyUI setting tests on the same RTX 4090.

Tested scope: one studio presenter portrait and one narration. Eight of the 24 settings combinations have captured evidence. Other portrait types and the remaining sixteen combinations are guidance or planned tests, not validated results.

Choose a starting preset for your goal

For a first draft, try crop · still OFF · 256 · enhancer none. Move to 512 for a larger face display, turn still ON for a steadier head, and compare GFPGAN before keeping its restored output. All seven presets below used batch size 2, expression scale 1.0 and pose style 0 in the captured evidence.

Swipe the table sideways to see all settings.

Measured settings on one studio presenter portrait
Goal	Preprocess	Still	Base size	Enhancer	Evidence
Quick preview	crop	OFF	256	none	Measured on the studio portrait
Larger face display	crop	OFF	512	none	Measured on the studio portrait
Optional face restoration	crop	OFF	512	gfpgan	Measured on the studio portrait
Steady presenter head	crop	ON	512	gfpgan	Three measured still-ON runs
Keep the scene, steady head	full	ON	512	gfpgan	One supplemental run only
Keep the scene, moving head	full	OFF	512	gfpgan	Three measured preprocessing runs
Already suitable face framing	resize	OFF	512	gfpgan	Three measured preprocessing runs

The steady-scene preset has only one supplemental render; it is a useful example, not repeat-trial validation. Full retains the surrounding source scene by compositing the animated face; it does not animate a whole body. GFPGAN ON is optional even for a final render.

What the measured results support

Still controls generated head motion, not scene framing

At crop, size 512 and GFPGAN ON, the OFF median was 117.40 s and ON was 117.68 s, from three measured runs each. ON kept this head nearer the source pose while the mouth animated. This test does not establish a speed advantage or guaranteed elimination of seams. Watch still ON/OFF.

Choose preprocessing for the composition

At size 512, still OFF and GFPGAN ON, full took 218.55 s, crop 118.13 s and resize 116.08 s (three-run medians). Full preserved the jacket and scene; crop made a face close-up; resize rendered the resized whole portrait. Full’s larger output also undergoes enhancement, so this does not isolate preprocessing cost. Compare the framing.

512 showed finer-looking detail, at greater cost in this sample

With crop, still OFF and GFPGAN OFF, 256 took 27.88 s and 512 took 55.41 s. Sampled whole-GPU peaks were 4,799 and 10,047 MiB. These are single trials, not medians. Final videos were 256 × 256 and 512 × 512. Inspect matching close crops.

GFPGAN adds generated detail, processing and final upscaling

At crop, size 512 and still OFF, OFF took 55.41 s and ON 119.39 s, one trial each. Sampled whole-GPU peaks were 10,047 and 10,207 MiB. ON produced 1024 × 1024 video and looked sharper around eyes and facial hair, with smoother skin. Sharper presentation does not prove recovered identity or improved lip sync. Compare restoration OFF/ON.

These are separate studies with different repeat counts. Keep their trial scopes intact rather than pooling timings into one ranking. Timing includes ComfyUI input loading, fresh worker/model startup, rendering and export; excludes queue wait and machine setup. VRAM is sampled whole-device usage, including a 396 MiB baseline in the quality study, and may miss brief peaks. The standalone GPU comparison uses a different execution path.

Adapt the starting point to the portrait

The factorial evidence is from one well-lit, front-facing studio portrait. The following choices for other inputs are guidance to try, not separately tested portrait presets.

Tight headshot: start with crop. If the source is already framed exactly as needed, inspect resize as an alternative; review crop boundaries before a long render.
Shoulder-up presenter or wider scene: try full when clothing and background must remain. Still ON may reduce distracting head movement; inspect jaw/neck seams. A small face in a wide image still limits available facial detail.
Soft or compressed photo: first compare an unenhanced draft. Try GFPGAN only if it looks useful; inspect the original likeness, skin and teeth. Restoration cannot establish the exact detail lost in the source.
Illustration or cartoon: try enhancer none first to avoid adding unwanted photorealistic texture. Face detection, motion quality and this recommendation were not benchmarked on cartoons here.
Profile, occluded, tiny or multiple faces: use a clearer single-face source where possible. No tested preset here establishes reliable handling or face selection for these cases.

Use the presets in ComfyUI or the standalone CLI

Follow our ComfyUI installation guide, then load a preset workflow, upload the shared portrait and narration, and Run. Every download below is a single-node preset adapted from the guide template. Loading a workflow does not launch a render by itself. An unchanged graph can reuse cached outputs.

Control mapping: ComfyUI still ON is the boolean true (shown as “still”); OFF is false (“animated”). The CLI uses --still for ON and omits it for OFF. ComfyUI enhancer none maps to omitting --enhancer in the CLI. Size is the base face renderer; the final video dimensions also depend on preprocessing and enhancement.

CLI examples specify the same settings; they were not timed as new standalone runs. Run them from the SadTalker repository in its activated environment, with the supplied inputs available. Use the four-thread launcher instructions to match the tested worker CPU limits; GUI workflows alone do not set these limits.

Quick preview

Measured on the studio portrait. Inspect its source test before adapting it to your own portrait.

ComfyUI workflow ComfyUI API graph

Quick preview · standalone settings example

bash

python inference.py --source_image "sadtalker-corporate-presenter-source.png" --driven_audio "sadtalker-male-narration-audio.wav" --result_dir "results/preview" --preprocess crop --size 256 --batch_size 2 --expression_scale 1.0 --pose_style 0

Larger face display

Measured on the studio portrait. Inspect its source test before adapting it to your own portrait.

ComfyUI workflow ComfyUI API graph

Larger face display · standalone settings example

bash

python inference.py --source_image "sadtalker-corporate-presenter-source.png" --driven_audio "sadtalker-male-narration-audio.wav" --result_dir "results/large-face" --preprocess crop --size 512 --batch_size 2 --expression_scale 1.0 --pose_style 0

Optional face restoration

Measured on the studio portrait. Inspect its source test before adapting it to your own portrait.

ComfyUI workflow ComfyUI API graph

Optional face restoration · standalone settings example

bash

python inference.py --source_image "sadtalker-corporate-presenter-source.png" --driven_audio "sadtalker-male-narration-audio.wav" --result_dir "results/restored-face" --preprocess crop --size 512 --batch_size 2 --expression_scale 1.0 --pose_style 0 --enhancer gfpgan

Steady presenter head

Three measured still-ON runs. Inspect its source test before adapting it to your own portrait.

ComfyUI workflow ComfyUI API graph

Steady presenter head · standalone settings example

bash

python inference.py --source_image "sadtalker-corporate-presenter-source.png" --driven_audio "sadtalker-male-narration-audio.wav" --result_dir "results/steady-presenter" --preprocess crop --size 512 --batch_size 2 --expression_scale 1.0 --pose_style 0 --still --enhancer gfpgan

Keep the scene, steady head

One supplemental run only. Inspect its source test before adapting it to your own portrait.

ComfyUI workflow ComfyUI API graph

Keep the scene, steady head · standalone settings example

bash

python inference.py --source_image "sadtalker-corporate-presenter-source.png" --driven_audio "sadtalker-male-narration-audio.wav" --result_dir "results/steady-scene" --preprocess full --size 512 --batch_size 2 --expression_scale 1.0 --pose_style 0 --still --enhancer gfpgan

Keep the scene, moving head

Three measured preprocessing runs. Inspect its source test before adapting it to your own portrait.

ComfyUI workflow ComfyUI API graph

Keep the scene, moving head · standalone settings example

bash

python inference.py --source_image "sadtalker-corporate-presenter-source.png" --driven_audio "sadtalker-male-narration-audio.wav" --result_dir "results/moving-scene" --preprocess full --size 512 --batch_size 2 --expression_scale 1.0 --pose_style 0 --enhancer gfpgan

Already suitable face framing

Three measured preprocessing runs. Inspect its source test before adapting it to your own portrait.

ComfyUI workflow ComfyUI API graph

Already suitable face framing · standalone settings example

bash

python inference.py --source_image "sadtalker-corporate-presenter-source.png" --driven_audio "sadtalker-male-narration-audio.wav" --result_dir "results/already-framed" --preprocess resize --size 512 --batch_size 2 --expression_scale 1.0 --pose_style 0 --enhancer gfpgan

The 512 and GFPGAN presets require their corresponding weights. Use the model guide, missing-checkpoint fix or FFmpeg checks if a render fails.

Inspect all four setting tests

Still ON vs OFF →

Head motion and presenter steadiness. Includes original videos, settings screenshots and downloadable evidence.

Full vs crop vs resize →

Scene framing and compositing. Includes original videos, settings screenshots and downloadable evidence.

256 vs 512 →

Base detail, render time and sampled VRAM. Includes original videos, settings screenshots and downloadable evidence.

GFPGAN ON vs OFF →

Restored detail, artifacts and processing cost. Includes original videos, settings screenshots and downloadable evidence.