SadTalker preprocess: full vs crop vs resize

Same input, three framing choices

The order is full → crop → resize. Only preprocess changes. All three use size 512, still OFF, GFPGAN, batch size 2, expression scale 1.0 and pose style 0. The synchronized comparison uses one audio track.

What changed visually?

full

The jacket and original surrounding frame remain visible. The animated face moves within that scene; inspect the jaw and neck transition as the head turns. The larger final image does not mean the face model rendered more detail than the other modes.

crop

The automatic crop produces a square close-up of the face, removing most shoulders and the wider source composition. Mouth, eyes and head movement remain visible. At the same display width the face appears larger, which also makes small mouth and edge artifacts easier to inspect.

resize

The output retains the wider portrait without pasting the face back into a static source frame. The face and nearby collar/shoulder area are rendered as part of that image. It completed on this portrait; that does not establish that resize will work well when a face occupies a much smaller part of the source.

Compare the jaw and neck boundary, face shape, teeth, eyes and background motion. These are visual observations from one portrait, not scored lip-sync or identity measurements. The template does not expose a seed, so expressions and blinking can vary between runs.

Measured results on the same RTX 4090

Each mode had one warm-up and three measured runs. We rotated the order between measured rounds: full/crop/resize, crop/resize/full, then resize/full/crop. Timing uses ComfyUI’s server execution events and includes input loading, worker/model startup, inference and video export; queue wait, installation and warm-ups are excluded.

Swipe the table sideways to compare all three modes.

Full took about 1.85× the crop median in this test. Crop and resize were close: their medians differ by 2.05 seconds. This setup combines full-image compositing with GFPGAN enhancement of a larger frame; it does not isolate the cost of preprocessing alone. Choose framing first, then check speed on your own inputs.

size = 512 is the base face-rendering setting, not a guarantee of a 512 × 512 final video. Full pastes the animated face into the source frame; the tested GFPGAN stage also upscales by 2×. Resize output dimensions can be affected by the source aspect ratio and video encoding. The table reports actual files, not inferred resolutions.

Reproduce it in ComfyUI

Use our ComfyUI setup guide to install the isolated SadTalker node. The comparison extends the guide’s LoadImage / LoadAudio template with three SadTalker nodes.

1. Get the three-mode workflow and drag the JSON into ComfyUI.
2. Upload the test portrait and the narration into LoadImage and LoadAudio. Both feed all three nodes.
3. Set preprocess to full, crop and resize respectively. Keep still OFF (the widget displays animated) and all other settings identical.
4. Click Run. Videos appear in the nodes and are saved under ComfyUI/output/sadtalker/<run-id>/. An unchanged graph may return cached outputs; fresh input filenames were used for the timing tests.

for mode in full crop resize; do
  python -c 'import json,sys; print(json.dumps({"prompt":json.load(open(sys.argv[1]))}))' "$mode-api.json" > prompt.json
  curl -X POST http://127.0.0.1:8188/prompt \
    -H 'Content-Type: application/json' \
    --data-binary @prompt.json
done

Which preprocess mode should you use?

Failure cases and practical limits

Observed no-face errors and limits

All twelve portrait renders produced video and audio successfully. A separate uniform gray image (1117 × 1034, no face) with a two-second narration excerpt failed in both full and crop. Both logs point to missing source landmarks and end with “TypeError: exceptions must derive from BaseException.” Further probe work was stopped at the owner’s request; no verified resize failure-probe result is included. These probes are excluded from the portrait medians.

The installed crop/full preprocessing requires detectable facial landmarks. If face detection fails, use a clear, front-facing portrait and inspect the worker log. Resize skips the automatic crop, but that does not make a missing or very small face usable. The two no-face errors are observed results; the resize advice is based on the implementation, not a tested success rate.

Get the no-face fixtureGet the two failure records

What this test cannot establish

• All twelve main portrait renders completed, but successful execution does not guarantee clean face edges, accurate lip sync or stable identity.
• One portrait, one narration and one GPU do not establish universal quality or speed rankings. Very small faces, profile poses, occlusion and multiple-face selection were not tested.
• GFPGAN was enabled for every mode. Enhancement can alter facial detail and adds processing, particularly when the resulting frame is larger.
• If you hit an out-of-memory error, try size 256 and batch size 1 for all three modes. Those settings were not benchmarked here.

Implementation references: input preprocessing, output framing and enhancement, and the tested ComfyUI node.