Higgsfield Troubleshooting Guide

QUICK FACTS

Generated-checked block (scripts/build_index.py verifies anchors). Read the linked sections for full context — these lines are routing aids, not the rules themselves.

• Face inconsistency, dead camera moves, ignored prompts, static i2v, blocked dark content — the per-problem fix list →
• Kling 3.0 Motion Control failures are almost always upstream of the prompt: reference clip, character image, or orientation/scene-source settings →
• Pre-generation checklist: subject, action, named camera preset, style, grade, aspect, <200 words (short-form regime) →
• Seedance/Cinema Studio symptom table + diagnostic flowchart: blurry = overspecified; chaotic camera = One-Move Rule violated; wrong character = prompt text contradicts the reference — delete the contradicting text first; delete ALL appearance text only on a start-frame / I2V shot (P1-1) →
• Every delivered take gets ONE of five verdicts before anything re-fires: keep / fix-in-post / edit / re-roll / rewrite →
• Two takes with the same flaw = rewrite, by rule; different flaws per roll = stochastic → batch-and-cull, not rewrite — against Retry Ladder rung 2 when both takes failed: OPEN, no default (§ Stop-Rule Ladder) →
• Re-roll = same prompt again, unchanged — no seed parameter on this surface; every roll is a fresh sample →
• Change exactly one variable between takes so causality stays readable →
• Declare the take budget AND a written "good enough" bar before take one; half-budget with no progress forces a strategy change →
• The shot log is the ledger row — one line per take, changed variable in notes →
• Continuation/extension defects: 12-row symptom → cause → single-repair-variable atlas (planned-vs-observed opening, motion-vector drop, prop contradictions, chain-depth drift…) →
• Retry Ladder: 4 terminating rungs — re-run once verbatim → treat 2nd failure as over-packing (when the two failures differ, Take Triage reads them as stochastic: OPEN, no default — § Stop-Rule Ladder) → switch model for that shot → stop after 3 paid attempts with named options →
• Six stop numbers, one ladder: 2 same-flaw re-rolls → rewrite · 3 paid attempts (no declared budget) → named options · half a declared budget with no progress → change strategy · 4 v2v batches (ceiling) → prompt/source fault · 10–15 surgical iterations (declared budget) → simplify the shot · 65–100 generations per kept shot is a project benchmark, never a stop rule. The earliest tripwire wins; where 2 and 4 count the same v2v runs it is OPEN, default the earlier stop →
• Log EVERY confirmed fix to learning memory, and check memory first before troubleshooting →
• Vision-grounded diagnosis (stills only): vision proposes the reject_reason, the human confirms — advisory until a class clears the agreement gate →