How do I avoid result as bounding box when using llama-qwen2vl-cli ?

[gnusupport]

My prompt is:

PROMPT="Describe image and extract full text if any, for this image named: $NAME."

and I run it this way:

OUTPUT=$(llama-qwen2vl-cli -n 8192 -c 8192 -m $MODEL_FILE \
  --mmproj $MMPROJ_FILE \
  --no-perf --image "$IMAGE_FILE" -p "$PROMPT" 2>/dev/null | \
  grep -v -e "clip_init:" -e "clip_ctx:" -e "encode_image_with_clip:" | \
  sed -e ':a;N;$!ba;s/^[[:space:]\n]*//;s/[[:space:]\n]*$//')

but I receive sometimes similar outputs as this one:

<|object_ref_start|>IMG_20250307_191036_344<|object_ref_end|><|box_start|>(568,444),(751,995)<|box_end|>

is there way to avoid getting this output? It happens with some pictures, though not all.