Controlling character pose with a reference image in SDXL: IP-Adapter vs ControlNet conflicts
Unfair-Walk-9805 · reddit · 2026-09-13
The author generates 128×128 pixel art poses of one character: preprocess reference images into pixel-art style, use IP-Adapter for appearance, ControlNet pose/rig for target pose, and experiment with multi-view references plus pose/depth annotations. The core issue is conditioning conflicts — the model follows ControlNet's pose but duplicates limbs from the reference image. Strength and timing tweaks help only somewhat; training per-character LoRAs is too costly. They ask for workflows that preserve appearance while controlling pose.
More from coding & agent
- Is the final UI clear? 'Language in, intelligence, language out' — and agents as the value layer — eptwts · 2026-09-13
- Dev shares 3 personal AI apps he uses daily: a 24/7 voice desktop agent and a GPUI Messages rewrite — johnlindquist · 2026-09-13
- Adopt one tip at a time: fix your current annoyance, then bake it into your instructions — TheMoonMidas · 2026-09-13
- Feeding your usage log to ChatGPT to compare Codex orchestrator setups on cost — TheMoonMidas · 2026-09-13
- Stale rules in agents.md may cause Codex's Astra to pause constantly — ask it to show you — TheMoonMidas · 2026-09-13
- Give Codex a clear stopping point: reversible details on its own, ask before publishing or spending — TheMoonMidas · 2026-09-13