Skip to main content

Image & Video Generation

Image and video diffusion​

Build-time flags define the compiled shape and denoising profile. Runtime flags select request values within the bundle's contract.

trtmc build MODEL_ID \
  --image-height 1024 \
  --image-width 1024 \
  --num-inference-steps 28 \
  -o diffusion.bundle

trtmc generate-video diffusion.bundle \
  --prompt "A sunrise over a mountain lake" \
  --output frames \
  --num-steps 28

Video models can additionally compile height, width, and frame count with --video-height, --video-width, and --video-num-frames. A runtime request must remain within the profiles packaged by the exact family build.

Classification and segmentation​

trtmc classify classifier.bundle --image input.jpg
trtmc segment segmenter.bundle --image input.jpg --output mask.png
trtmc segment-prompted prompted.bundle \
  --image input.jpg --output masks --point-x 0.5 --point-y 0.5

The public CLI also reserves detect, but an API command alone is not a supported-model claim. Confirm an exact model-owned strategy and E2E manifest in Model Recipes, organized by the exact Hugging Face image or video task.

For the denoising-loop and task-result mental model, follow the Diffusion, Vision, and Time-Series Tutorial.