OpenShotDownloadAndLoadSAM2Model
OpenShotSam2Segmentation (single-image)
OpenShotSam2VideoSegmentationAddPoints
OpenShotSam2VideoSegmentationChunked (meta-batch/chunk friendly)
OpenShotGroundingDinoDetect (text-prompted object detection -> mask + JSON)
OpenShotTransNetSceneDetect (direct TransNetV2 inference -> IN/OUT JSON ranges)

Why this exists

Some existing SAM2 video nodes accumulate entire-video masks in memory and can OOM on longer clips. This pack provides chunk-safe behavior while keeping Comfy-native node patterns.

Attribution

This project is inspired by and partially based on ideas and APIs from:

kijai/ComfyUI-segment-anything-2
Meta SAM2 research/code

Please see upstream projects for full original implementations and credits.

Requirements

ComfyUI
PyTorch (as used by your Comfy install)
decord (required by many SAM2 video predictor builds)
sam2 Python package/module available in your Comfy runtime
GroundingDINO runtime via Hugging Face Transformers (installed from requirements.txt)
transnetv2-pytorch (installed from requirements.txt)

Install this node pack into ComfyUI/custom_nodes/OpenShot-ComfyUI and restart ComfyUI.

Install dependencies

/home/jonathan/miniforge3/envs/comfyui/bin/python -m pip install -r /home/jonathan/apps/ComfyUI/custom_nodes/OpenShot-ComfyUI/requirements.txt

Important note about SAM2 installs

requirements.txt intentionally does not install SAM2 from Git. This avoids repeated large temporary downloads and long install times during routine updates.

Ensure your Comfy environment already has a compatible sam2 installed.

GroundingDINO model downloads

OpenShotGroundingDinoDetect will download selected model weights from Hugging Face on first use and cache them in your HF cache directory.

TransNet scene detection behavior

OpenShotTransNetSceneDetect runs TransNetV2 inference directly and returns scene range JSON (start_seconds, end_seconds), so no external TransNet Comfy node pack is required.

Models

OpenShotDownloadAndLoadSAM2Model auto-downloads supported SAM2 checkpoints into ComfyUI/models/sam2 if missing.

Notes

OpenShotSam2VideoSegmentationChunked returns only the requested chunk range (bounded memory) instead of collecting whole-video masks.
For very long videos, pair chunked outputs with batch-safe downstream nodes (VHS meta-batch, staged processing, or on-disk intermediates).