PEOPLE, architectural detail, automatic SAM2 and the RGB atlas
The full graph contains several independent mask systems. They should not be collapsed into one abstract “mask”: each has its own source, coordinate space and downstream consumer.
Every mask has its own source, polarity and consumer
Source and purpose of each mask family
| Family | Route | Purpose | Main consumers |
|---|---|---|---|
| PEOPLE | 550 → 114 → 115 → 144 → 146 | Segment person/human/face/hands/feet/accessories | 420, 500 |
| Architectural detail | 591 → 580 → 584 → 585 | Segment building/facade/detail | 583, 775 |
| Automatic SAM2 | 234 + 779 → 232 → 233 | Automask diagnostic | Preview 233 |
| RGB/CMY/B/W atlas | 338 → 337 → 339/346/347/348/349/350/352/354 | Split a color-coded mask image | Previews 340–353 |
Florence2 coordinates drive the SAM2 single-image model
Florence2Run 550 receives resized source 780 and model 112. Coordinates 114 feed SAM2 115 together with single-image model 107. GrowMask 144 expands the region by 5 px; MaskBlur+ 146 creates a soft 10/auto edge.
Runtime testing confirmed that detection and SAM2 must receive the same image/canvas. For multiple Florence2 BBOX results, the local Sam2Segmentation version requires individual_objects = ON; with OFF, the test retained only one silhouette.
Multi-person runtime
On a 1280 × 720 source, indices 0–5 with individual_objects ON produced a combined mask for six people.
Production composite
Runtime testing confirms the mask chain separately; downstream positioning and compositing are validated by the next probes.
The detail mask is built on selected base image 592
Prompt 591 and source 592 feed Florence2Run 580. Coordinates 584 and SAM2 585 create the mask used by MaskToImage 583 and ImageCompositeMasked 775.
MASK↔IMAGE conversions require visual verification
No explicit InvertMask
There is no separate active InvertMask in the significant topology between PEOPLE detection and composite.
Hidden inversion
Third-party nodes may interpret intensity/polarity internally; this requires preview or runtime testing.
QA rule
Check the white foreground region, black background, holes, islands and dimensional match with the image.