Artur Gonçalves

**Is your feature request related to a problem? Please describe.** The program could be faster by outputting the frames directly to ffmpeg, via a pipe, instead of saving PNG files...

Off-by-one error was causing the undistorted image to be stretched horizontally and missing 16 pixels on the right side. Also, for some reason, `meshCellSize=16` causes glitches for me, but `20`...

ggml is compiled with cublas, but GPU is not used

2

I compiled ggml with `-DGGML_CUBLAS=ON` and then clip.cpp, and used it to get text encodings, but the GPU is not being used. The code takes the same amount of time...

Same fix as [here](https://github.com/AILab-CVC/YOLO-World/issues/466#issuecomment-2288000037) from @shiboyang Tested with ```sh python deploy/export_onnx.py \ configs/pretrain/yolo_world_v2_l_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py \ demo/pretrained_weights/yolo_world_v2_l_obj365v1_goldg_pretrain-a82b1fe3.pth \ --custom-text custom-text.json \ --opset 12 --without-nms python deploy/onnx_demo.py \ work_dirs/yolo_world_v2_l_obj365v1_goldg_pretrain-a82b1fe3.onnx \ custom-text.json \ --onnx-nms...