AD_sam_Crop

"Describe the object, and I'll track it through the whole clip, cropping it into a stable, same-size box on every frame." That's the pitch for AD_sam_Crop, and it's the first node in the pack's track-and-edit pipeline. It runs SAM3 video object tracking over your footage, then produces two things the rest of the pipeline needs: a stack of fixed-size per-frame crops (crop_img) and the tracking metadata (transform) that lets the downstream nodes paste edits back into the right place.

It sits at the front of a pipeline that looks like this: AD_sam_Crop → (feed the crops to AD_Inject_Latent to re-generate that region) → AD_sam_stitch (paste the refined crops back). If you've ever re-generated a face in a video and watched it swim, it's because the crops weren't stabilized - this node fixes that by smoothing the track before it ever reaches the sampler.

The inputs that matter

• pos (STRING) - the object description. This is the key beginner input and the one people get wrong. It's a text prompt describing the object to track (e.g. "a man in a red jacket"), which gets CLIP-encoded into the conditioning that drives SAM3. Empty pos = instant failure.
• image (IMAGE) - the frames, e.g. the IMAGE output of AD_In_VideoSplit.
• ckpt_name - the SAM3 checkpoint, default sam3.1_multiplex_fp16.safetensors. It loads through CheckpointLoaderSimple, so the file goes in ComfyUI/models/checkpoints, not a SAM-specific folder. The model+prompt pair is cached, so re-runs with the same settings are fast.
• detection_threshold (FLOAT, default 0.5) - how confident SAM3 must be to accept a detection. Lower it when tracking is missing frames.
• max_objects (INT, default 4) - caps how many object instances are tracked.
• detect_interval (INT, default 1) - how many frames between detections (the track is interpolated between them).
• crop_factor (FLOAT, default 3) - how much padding around the object the crop keeps. 3 = the crop is 3× the object's bounding box. More context for the sampler, more chance of drift.
• crop_width / crop_height (default 512×384) - the fixed crop canvas size. Must match the generation resolution on the H3 side - this is the #1 wiring mistake.
• smoothing_preset - the seven presets (balanced, stable_max, stable_extreme, cinematic_push, responsive, static_shot, cg_animation) pick the smoothing window and method (Gaussian or Savitzky-Golay) for the track. The tooltips are genuinely useful: stable_extreme for shakycam/low-bitrate footage, responsive for fast handheld moves, static_shot for a locked tripod, cg_animation for clean renders that need almost no smoothing. balanced really does handle ~80% of footage.

Outputs: crop_img (IMAGE), transform (H3FACEXFORM), track_data (SAM3_TRACK_DATA, the raw track for other SAM3 nodes), masks (MASK - the per-frame object masks).

How it works

It runs ComfyUI's native SAM3_VideoTrack + SAM3_TrackToMask, derives a bounding box per frame from the masks, then interpolates and smooths the box centers and sizes over time (Gaussian or Savitzky-Golay per the preset) so the crop doesn't jitter. From the smoothed boxes it builds an affine transform per frame - that's the transform you pass around - and crops everything to the fixed canvas. Unstable tracks get their confidence recorded as per-frame weights, which downstream nodes use to fade out bad frames instead of glitching.

Common failure modes

The big one: "SAM3 did not detect the requested object in any frame. Lower detection_threshold or change pos." - that's the node telling you the description didn't match anything, or the threshold was too strict. Also note this uses ComfyUI's native SAM3 (comfy_extras.nodes_sam3), not the third-party ComfyUI-SAM3 pack - which in 2026 had a genuinely bad reputation for installing its own dependencies and breaking other node packs. Don't install that one for this; you only need the native support plus the checkpoint.

Installing it

From cardenluo/ComfyUI-Apt_Preset:

cd ComfyUI/custom_nodes
git clone https://github.com/cardenluo/ComfyUI-Apt_Preset
pip install -r requirements.txt   # or install.bat

Then drop sam3.1_multiplex_fp16.safetensors into ComfyUI/models/checkpoints and restart. You need a ComfyUI build recent enough to ship native SAM3 nodes.