hermes-skills/ltx-video-pipeline/references/artifact-deep-research-template.md

Artifact Deep Research — Question Template

Template for dispatching a focused deep-research pass when LTX-2.3 artifacts persist after applying the ranked fixes. Used 2026-07-22 for Story 1 "Ghost in the Wire."

Structure

1. OUR SETUP — exact model chain, render settings, hardware
2. STORY PROMPTS — all scene prompts verbatim
3. WHAT WE'VE ALREADY TRIED — v4 and v5 settings + results
4. EXISTING RESEARCH — path to prior artifact research
5. WHAT WE NEED NOW — 10 specific research questions targeting our exact setup

Dispatch Command

hermes -p research -s deep-web-research chat -q "Read the full research question from /home/n8n/workspace/research/<date>-ltx-artifacts-question.md and execute it exhaustively..." -Q --max-turns 600 --yolo

Key Research Questions (from 2026-07-22 dispatch)

1. Given our EXACT prompts, what specific artifacts would LTX-2.3 produce with these scene descriptions?
2. Does ID LoRA (talkvid-3k, trained for talking-head) cause artifacts on full-body action scenes?
3. Is the LTX Director refiner stage (4 steps, denoise 0.42) introducing artifacts?
4. Is fp8_scaled the worst fp8 variant? Compare vs fp8_e4m3fn vs v1.1 for I2V with LoRAs.
5. Does 512×512 square aspect produce more artifacts than widescreen?
6. Is Distilled LoRA at 0.7 still too high when stacked with ID LoRA?
7. Are there NEW fixes or LoRAs released in the last 30 days?
8. Could Flux.1-dev start frame inconsistencies cause artifacts?
9. Are there known LTX Director node bugs beyond the wiring bug we already fixed?
10. Does LTX-2.3 struggle with cyberpunk/neon-heavy scenes specifically?

Expected Output

/home/n8n/workspace/research/results/<YYYY-MM-DD>-ltx-2.3-artifacts-deep.md with YAML frontmatter and findings ranked by likelihood of causing OUR specific artifacts.