Hand-drawn line diagram as 2nd ControlNet fixes archer's grip but not bowstring
Sensitive-Wealth5801 · reddit · 2026-09-08
A detailed SDXL experiment: feeding a hand-drawn line diagram to a second ControlNet (Union model, SCRIBBLE mode) stacked on OpenPose fixed the archer's grip, but bowstring and arrow topology still fail.
Pipeline: SDXL + CHEYENNEv20, dpmpp2m/karras, 34 steps, CFG 6.0, 832x1216; CN1 OpenPose at strength 1.0, endpercent 0.75; CN2 scribble-mode line diagram at strengths 0.40/0.50, 3 seeds each.
Results: grip and anchor pose now correct, but the string cuts diagonally without attaching to the lower limb, three shafts appear instead of one, and the arrow crosses in front of the bow.
Ruled out: text descriptions (FLUX.2 img2img also failed on the string), strength variations (identical defects at 0.40 and 0.50), and stroke-width asymmetry (9px limbs vs 6px string) as sole cause.
Questions for the community: minimum stroke width for isolated lines in scribble inputs; whether lineart/canny/mlsd hold thin straight lines better; proper strength/startpercent for a second stacked ControlNet; forcing exactly one arrow; whether regional prompting/attention couple or weapon-region inpainting works better.
More from coding & agent
- What broke running an agent fleet 24/7 — it wasn't prompt quality — ParrotIntegrated · 2026-09-08
- Microsoft shipped Friday; by Tuesday a stranger had reverse-engineered it with official tokens — unixterminal · 2026-09-08
- MaintainabilityBench: grade AI on the cost of adding features, not correctness — kuza55 · 2026-09-08
- Utilify MCP connector lets AI agents compare and sign up for Texas utility plans — modelcontextprotocol · 2026-09-08
- Astra's experimental compaction is a global setting that hits subagents — brandon_galang · 2026-09-08
- Dev texts random SF track ideas to Astra agent, which ships each one — blakesamic · 2026-09-08