18d • Help
Quick tip of the week: Lower your Hires denoise to get a cleaner LoRA result (SDXL)
If your character looks perfect in the preview, but the final image changes the face… it’s usually not your LoRA.
Most of the time it’s Hires.fix / Upscale denoise re-drawing the face during the second pass.
✅ What to do
- Keep your LoRA weight the same (ex: 0.35–0.55)
- Lower Hires denoise so the upscale doesn’t “recreate” the identity
🎯 My go-to SDXL settings
- Upscale: 1.3x – 1.7x
- Hires steps: 10–15
- Hires denoise: 0.20 – 0.35 (this is the magic range)
- If you go 0.45+, expect face drift and “AI beauty” artifacts
Why it works:
Low denoise = the upscale keeps your existing face/details.
High denoise = it invents new features and your character “drifts”.
4
0 comments
Quick tip of the week: Lower your Hires denoise to get a cleaner LoRA result (SDXL)
powered by
skool.com/ai-influencer-model-training-4492
Learn Stable Diffusion for hyper-real NSFW AI Models and monetize them on Fanvue.
Step-by-step video tutorials from zero to pro. Join!
Suggested communities
Powered by