Skip to main content
Split Parts

Split Parts · screenshot coming soon

​
What this node does

Split Parts runs a segmentation workflow (backed by a SAM-style model) to cut an image into individual parts. You tell it what to isolate in one of two ways:
  • On the card, click points on the parts you want (positive/negative point prompts) or draw boxes around them. Those prompts are stored as JSON in parts_data.
  • By text (with a text-driven workflow), type a concept in main_prompt such as strap, buckle, or logo.
Each point group or box becomes a separate segmentation call, and the resulting part images are merged into one numbered batch. This is a ▶ Run stage: it dispatches the chosen workflow to a Runner (local ComfyUI or a remote server) and needs a GPU pass, so it is not instant. After a successful run the parts appear as a numbered grid on the node; clicking a thumbnail sets selected_index to pick one for the single-image output. Outputs are ComfyTV snapshots, not native tensors — insert a Bridge to feed native ComfyUI nodes (bridges docs).

​
When to use it

  • Break a product photo into components (buckle, strap, sole) to edit or re-texture each separately.
  • Isolate a logo or badge from a garment for cleanup or replacement.
  • Pull several named objects out of one frame in a single pass, then pick the one you need downstream.

​
Parameters

​
workflow

Which segmentation workflow to run (dropdown, options from the split-part workflow set, with a default). Text-prompt-driven workflows use main_prompt; point/box workflows use what you draw on the card.

​
main_prompt

The concept to segment when using a text workflow — e.g. strap, buckle, logo. If you neither drew points/boxes nor typed a prompt, the stage errors and asks you to do one of them.

​
image

Optional COMFYTV_IMAGE input — the source image to split. Connect the picture whose parts you want to separate.

​
parts_data

Internal (hidden). JSON of the point groups and boxes you drew on the card, in source-image pixel coordinates. You never edit this by hand; the card UI writes it.

​
selected_index

Internal (hidden), default 1, 1-indexed. Which part from the batch feeds the single-image output. Set by clicking a thumbnail in the node’s output grid.

​
custom_params

Internal (hidden). JSON of any user-defined parameter attachments for this node.

​
Outputs

​
Tips

  • Multiple point groups / boxes each run as a separate segmentation, so you can pull out several parts in one Run; they come back merged and renumbered.
  • Positive vs negative points matter: negative clicks tell the model what to exclude, which helps carve one part cleanly away from a neighboring one.
  • If segmentation runs but returns nothing, the stage reports an empty result — try different points/boxes or a lower threshold in the workflow’s params.
  • Mask Cleanup — tidy the alpha of a part cutout before compositing.
  • Image Picker — an alternate way to pick one image out of a batch downstream.
  • Cutout — subject-vs-background separation when you want one foreground, not per-part pieces.