AV ComfyUI Manual
48 / 64
Files
48
PEOPLE / PPL · SPATIAL CONTRACT

Where person size, crop and coordinate space change

Generation and masking can be correct while the figure still disappears or lands in the wrong place if source image, mask bounds and destination canvas diverge across nodes 829, 780, 420, 449 and 429.

CONFIRMEDResize/crop/paste topology is confirmed; exact visual placement requires runtime imagery
COORDINATE SPACE · PEOPLE INSERT

Size, crop and canvas must align before Paste By Mask

827 / 829PPL canvas1312 × 1920
552People sourcelinked from 543
780Resize1920 × 1920
420 / 781Crop regionmask bounds
449 / 429Cut + Pastekeep_ratio_fit
01same image/mask size
02correct mask bounds
03person inside canvas
04original/resized return
01 · COORDINATE SPACES

PPL latent, detection canvas and scene canvas do not share one size

StageNodesStored / derived spaceRisk
PPL generation827 → 828 → 8291312 × 1920, batch 1Figure framing differs from the scene
Detection input552 → 780Resize to 1920 × 1920Aspect ratio / padding changes
Mask crop146 + 780 → 420 → 781Bounds from people maskLoose/tight crop or empty region
Prepared cutout422 → 477 → 449Cropped person + alphaImage/mask mismatch
First paste779 + 449 + 451 → 429Selected SDXL scene canvasScale/location outside frame
Alternative return509 → 684 → 782 → 685Restored original/resized canvasWrong mode 693
Runtime mask testFlorence2 → BBOX → SAM2Same 1280 × 720 source at both stagesDifferent images reproduce displaced/false masks
02 · FIRST PASTE

Node 429 connects the base scene, prepared cutout and position source

Paste By Mask 429 receives the base image from Image Filter 779, prepared person 449 and image/mask representation 451. Stored mode keep_ratio_fit affects the scale of the inserted person.

Output 429 passes through Images to RGB 672 and enters image1 of selector 459 for PPL mode 1.

CONFIRMED

Topology

429 → 672 → image1 of 459.

NOT CONFIRMED

Exact placement

Coordinates and figure visibility cannot be proven without runtime preview 480/459.

03 · QA ORDER

Verify positioning before looking for downstream errors

  • Before selecting a BBOX, prove that Florence2 and SAM2 receive the same image at the same size.
  • Compare actual dimensions of image 829 and resized image 780.
  • Verify that 146 contains foreground with the correct polarity.
  • Verify bounds/crop 420→781: the person is fully inside the region.
  • Verify alpha from 422/430 and alignment with RGB 477/449.
  • Verify first composite 429/672 before selector 459.
  • Verify mode 693 when using alternative return 685.
04 · INTEGRATION QA

A technically valid signal is not automatically a convincing composite

INFERRED

Perspective

Figure height and scale should match the horizon and architectural reference dimensions.

INFERRED

Grounding

Check foot contact, contact shadow and absence of floating.

INFERRED

Light / occlusion

Light direction, contrast and occlusions should agree with the environment.