Skip to main content

Image & Video

Image and video generation​

Build-time dimensions and frame count shape the TensorRT plans. Request-time options must remain within the owning family's contract.

python -m tensorrt_model_connect build MODEL_ID \
  --image-height 1024 \
  --image-width 1024 \
  --video-num-frames 49 \
  --output media.bundle

trtmc generate-video media.bundle \
  --runtime-root /opt/trtmc/lib \
  --prompt "A sunrise over a mountain lake" \
  --output frames \
  --num-steps 28 \
  --seed 7

Use generate-image for one image and generate-image-batch with a prompt file plus comma-separated seeds for a family that declares the batch Task.

Perception​

trtmc classify classifier.bundle \
  --runtime-root /opt/trtmc/lib --image input.jpg

trtmc segment segmenter.bundle \
  --runtime-root /opt/trtmc/lib --image input.jpg

trtmc segment-prompted prompted.bundle \
  --runtime-root /opt/trtmc/lib \
  --image input.jpg --point-x 0.5 --point-y 0.5 --foreground true

trtmc geometry moge.bundle \
  --runtime-root /opt/trtmc/lib \
  --image input.jpg --output geometry-output

Other current Task commands include extract-features, stereo disparity, and video-segment. Exact input shape, fixed profile, output representation, and supported checkpoint remain family-owned.

Continue with the Diffusion and Time-Series Tutorial.