Troubleshooting SDXL LoRA Training: Fixing 'Shiny Eyeball' Artifacts
MortytheMort · reddit · 2026-08-01
A developer using Ostris AI-Toolkit to train character LoRAs for SDXL reached out for help regarding a persistent issue.
- Setup: Running on a 12GB GPU, training takes about 1-1.5 hours with caching enabled and first sample skipped.
- The Issue: The trained LoRA is producing super bright, translucent, and shiny looking eyeballs.
- Attempted Fixes: The 30-image dataset includes various framings and lighting conditions, some with high exposure and catchlights. The author tried tagging these attributes in the dataset to prevent baking them into the checkpoint, but had no luck.
The author is looking for insights on whether this is caused by the dataset, training parameters, or other factors.
More from Multimodal
- Testing Dreamina Seedance 2.5: Seamless Lip-syncing with Stitched 30s AI Videos — churchkey · 2026-08-01
- Dreamina Seedance 2.5 Released: Native 30-Second Video Generation — PrajwalTomar_ · 2026-08-01
- Gemini-3 Stunning Demo: One Prompt Turns Any Location into a Time Machine — josh_bickett · 2026-08-01
- From Prompt to Physical: A ChatGPT to 3D Printing Workflow — Squigels · 2026-08-01
- Seedance 2.5 Video Generation Previewed, Coming Soon to Higgsfield — Aiden_Tech_Ai · 2026-08-01
- Seedance 2.5 Test: Generating Coherent AI Shorts with 42 Reference Images — DavidmComfort · 2026-08-01