FOMM Seg15 Chooser

FOMM_Seg15Chooser is the fine-grained end of the segment-picker trio in this pack. It generates the chosen_seg_indices string for FOMM_Partswap, matching the vox-15segments model - the one that splits a face into sixteen semantic regions so you can swap individual parts with real precision.

Its checkboxes are the most granular set in the pack: m_0_background, m_1_rt_fr_head, m_2_eyes, m_3_mouth, m_4_lwr_shoulder, m_5_shoulder, m_6_lt_fr_head, m_8_hair_tip, m_9_frnt_face, m_10_rt_ear, m_11_outer_hair, m_12_lt_ear, m_13_lt_hair, m_14_rt_hair, m_15_neck - and one quirky one, m_7, which the pack's constants list with no name at all. It's a real, detectable region; the authors just never labelled it. If you see a stray swapped patch, that's it.

Background defaults off, everything else on. As with the other choosers, this node does no inference: it reads the booleans, keeps the indices of the checked ones, and outputs them as an underscore-joined string for FOMM_Partswap to deserialize.

Why 15 segments

Precision, and a specific pairing. With 15 segments you can do things the 5- and 10-segment models can't express, like swapping just the mouth (m_3_mouth) for a lip-sync effect, or swapping the face while leaving every strand of hair (m_13_lt_hair, m_14_rt_hair, m_8_hair_tip, m_11_outer_hair) untouched. And per the README, the 15-segment model is the one the optional face-parsing pass (use_face_parser on FOMM_Partswap, which needs the resnet18-5c106cde.pth + 79999_iter.pth weights in face_parsing/) is designed to pair with for cleaning up residual background. The finer segments plus face parsing is the setup the author points at for the cleanest swaps.

Wiring it up

• Output chosen_seg_indices → FOMM_Partswap's chosen_seg_indices input.
• This chooser is for vox-15segments only. Use it with the 5- or 10-segment models and the index numbers stop meaning what they say - out-of-range indices get silently dropped, so you'll usually end up with a partial or empty swap instead of a crash.

Install and gotchas

Pack install only - ComfyUI Manager (search "FirstOrderMM") or git clone + pip install -r requirements.txt into custom_nodes/, restart. The model it feeds (vox-15segments.pth.tar in checkpoints/) is a manual download like everything here.

Two things to remember. First, the default (everything but background on) is a full-face swap; untick what you want to keep. Second, if you enable use_face_parser on the Partswap node, that's extra weights you have to fetch by hand - the README's "should definitely be used for FOMM models" refers to using it, but it won't do anything until the two .pth files are in place. And if your output has a mysterious unswapped or oddly-swapped patch around the forehead, you've likely met m_7 - the nameless segment.