58
FAILURE TRACE
Find the stage where people disappear
Diagnosis proceeds from upstream to downstream: do not tune a late composite until generation and masking are proven to work.
INFERREDThe diagnostic order is derived from confirmed preview / comparer nodes
FIVE-PROBE TEST
The last correct preview determines the next check
1
409Generateraw PPL decode2
113DetectFlorence2 preview3
507 / 508Cropsource + mask4
480CompositeMASK / PPL5
730Current finaldecode 53If stage N is correct, do not return to the prompt; inspect the connection and parameters of stage N+1.
Checkpoints already present in the workflow
| What to verify | Node | Interpretation |
|---|---|---|
| Generated people | 409 ← 829 | If empty, the problem is upstream of the mask chain |
| Florence result | 113 ← 550 | Detection should find the intended objects |
| Cut / selected crop | 507 ← 503; 508 ← 522 | Verify source switch and crop |
| Mask vs PPL | 480 ← 451 / 459 | Compare placement mask and selected composite |
| Inpaint vs source | 518 ← 509 / 552 | Compare processed result and selected source |
| Current final save | 730 ← 53 | Active result point of the saved configuration |
| Optional FLUX vs Upscale | 141 ← 53 / 833 | HQ / upscale chain is currently stored in bypass |
| Optional final preview | 15 ← 849 | Check only after enabling the overlay chain |
Symptom → likely failure area
| Last working stage | Most likely failure | Next test |
|---|---|---|
| No people in 409 | Prompt / model / sampler | Check 408 → 831 → 828 → 829 |
| Florence shows every box, SAM2 leaves one person | Sam2Segmentation multi-BBOX mode | Enable individual_objects and repeat with the same indices |
| Mask lands on a building, corner or empty region | Florence2 and SAM2 received different images/canvases | Feed the same-size source into detection and segmentation |
| 409 works, mask is wrong | 550 / 114 / 115 / 144 / 146 | Simplify classes and inspect the edge |
| Mask is good, cutout is dirty | 420 / 422 / 430 | Inspect crop and RemBg outputs |
| Composite exists, 459 selects the wrong result | 543 / linked selector inputs | Verify effective value = 1 |
| 730 contains people, final 15 loses them after HQ is enabled | Upscale / overlay / late overwrite | Compare 141, 730 and 15 |
What is known about the current case
Topology
The PPL branch reaches the final output; there is no broken link after node 459.
Likely area
If generation and mask are visibly present, investigate positioning, composite, selector output or a later overwrite.
Exact failure node
Without a current set of runtime previews, naming one guilty node would not be justified.
Diagnostic ladder from input to optional HQ
| Stage | Probe | Stop condition |
|---|---|---|
| Preprocessors | 39 depth / 167 edge | Map is wrong — do not begin SDXL diagnosis |
| Main SDXL | 71: input 79 vs result 779 | SDXL result is wrong — do not diagnose PEOPLE yet |
| Masks | 113 / 233 / 340–353 / 581 | Verify polarity, bounds and canvas size |
| PEOPLE | 409 / 507 / 508 / 480 / 518 | Find the last correct PPL stage |
| Main FLUX | 72: 779 vs 53; active save 730 | Verify survival after denoise 0.18 |
| Upscale / overlay | 141 / 675 / 15 / 531 / 293 | Remove bypass first, then verify |