test_temporal_offset/make_grid.sh

#!/usr/bin/env bash
# Stack the five side-by-side arms into one labelled comparison video.
# Each row is [pose reference | generated output], captioned with the RoPE shift applied, the lag
# it predicts, and the lag actually measured by score_arms.py.
set -euo pipefail

HERE=/home/xinyuy/flexvideo/LTX-2/test_temporal_offset
SRC=$HERE/out_viz/with_condition
FONT=/usr/share/fonts/truetype/dejavu/DejaVuSansMono.ttf

LABELS=(
  "A_base|A  reference | output   RoPE +0    predicted lag  0   measured  0"
  "B_rope_p8|B  reference | output   RoPE +8    predicted lag -8   measured -8"
  "C_rope_p16|C  reference | output   RoPE +16   predicted lag -16  measured -16"
  "D_rope_m16|D  reference | output   RoPE -16   predicted lag +16  measured +16"
  "E_content_p16|E  reference | output   CONTENT +16 (control)  predicted lag -16  measured -16"
)

inputs=() ; filters=() ; labels=""
i=0
for entry in "${LABELS[@]}"; do
  IFS='|' read -r arm text <<<"$entry"
  inputs+=(-i "$SRC/$arm.mp4")
  # Escape the caption for drawtext, then reserve a 40px black bar above each row for it.
  esc=${text//:/\\:}
  filters+=("[$i:v]scale=960:320,pad=960:360:0:40:black,drawtext=fontfile=$FONT:text='$esc':fontcolor=white:fontsize=17:x=8:y=12[v$i]")
  labels+="[v$i]"
  i=$((i+1))
done

ffmpeg -y -loglevel error "${inputs[@]}" \
  -filter_complex "$(IFS=';'; echo "${filters[*]}");${labels}vstack=inputs=$i[out]" \
  -map "[out]" -r 24 -c:v libx264 -crf 18 -pix_fmt yuv420p "$HERE/comparison_grid.mp4"

echo "wrote $HERE/comparison_grid.mp4"
ffprobe -v error -select_streams v:0 -show_entries stream=width,height,nb_frames -of csv=p=0 "$HERE/comparison_grid.mp4"