By name: ltx-fidelity-debugging description: Use this skill when debugging real LTX-2.3 output quality, running safe manual smoke tests, interpreting stage snapshots, or deciding whether a result is specific enough to count as a clear dog or other visually coherent subject.
LTX Fidelity Debugging
Use this skill for the narrow slice where the real LTX-2.3 path runs but the output is still visually wrong, ambiguous, or not strong enough to promote into repo truth.
This skill is for:
- real-weight local smoke runs
- stage-local visual debugging
- safe iteration on the
22BAV path - deciding whether the next suspect is prompt/stage 1, x2/stage 2, or decode/export
Do not use this skill for the common implementation path. The default repo loop still lives in AGENTS.md, MEMORY.md, and scripts/dev.py.
Default workflow
- Read
docs/research/10-ltx-fidelity-debug-playbook.md. - Use
scripts/ltx_debug_smoke.pywith a fixed seed and real converted artifacts. - Start with the named profile ladder:
safe-smoke, thenvisual-gate, thencoherence. - Keep clean lifecycle on unless you are disproving the lifecycle itself.
- Use detached
tmuxexecution for long real-weight smokes. - Inspect
stage1,post_x2, andfinalsnapshots before deciding what to fix next. - Review
frame_0001,frame_mid, andframe_lastbefore promoting a rung. - Downgrade the claim if the result is ambiguous.
What to load
docs/research/10-ltx-fidelity-debug-playbook.mddocs/research/07-ltx-integration-seams.mddocs/research/09-open-questions-and-validation-plan.mdMEMORY.md
Load these code paths only when you need the implementation seam:
packages/families/ltx/src/ltx/_prompt_encoding_backend.pypackages/families/ltx/src/ltx/_generation_backend.pyscripts/ltx_debug_smoke.py
Rules
- Start at the smallest truthful rung, not the highest resolution.
- Use real converted artifacts for visual conclusions.
- Change one major variable at a time.
- Trust stage-local evidence over the final clip when localizing a bug.
- Generic quadruped output does not count as dog success.