SadTalker quality and GFPGAN benchmark — 4 October 2026
Four successful fresh ComfyUI renders, same image/audio bytes, RTX 4090.
See benchmark.json for times, sampled whole-device VRAM, hashes and versions.
raw-evidence contains actual submitted API graphs (unique input names), native execution histories, original MP4s and worker logs.
assets contains stable reusable API/GUI workflows, matching display comparisons, close crops, captions and settings screenshots.
inputs contains the common original portrait and narration. Upload these files to ComfyUI/input before loading a reusable workflow.
One trial per configuration; no additional warm-ups or retries. No seed control. OFF and ON are independent SadTalker generations; GFPGAN effects are not isolated on identical generated frames.
VRAM is sampled nvidia-smi memory.used at 0.5s target plus query latency. Includes ComfyUI, worker, driver and other allocations; may miss brief peaks. Not minimum card requirements.
Comparison panels use bicubic interpolation at equal 512px display size; close crops approximate x24%, y35%, width/height52% (pixel rounding occurs).
Reproduction: use the existing ComfyUI guide and Python3.10 worker, tested revisions in report. benchmark.py temporarily installs a four-thread worker wrapper and restores original config in finally. Inspect setup paths before running elsewhere. limit_worker.py is a reusable thread-limits reference. Original GPU remains running; this archive does not manage rental billing.
