Trained an SDXL LoRA for Chinese gongbi portraits, then abandoned it
xiaoxuanchuang · reddit · 2026-11-03
A developer building a side project that turns selfies into Chinese court paintings trained his own gongbi-style SDXL LoRA to avoid API costs — then abandoned it. Setup: SDXL base 1.0 with kohya sd-scripts; faces cropped from paintings (Grounding DINO detection, manually fixed); v3 at its 2nd checkpoint won; inference was img2img on selfies at 640×640, 25 steps, cfg 7.5, strength 0.2, LoRA fused at 0.8. Why it failed: (1) the downstream InsightFace (buffalol) step often found no face — the better the gongbi style, the less the detector saw a face; (2) even when detected, identity was mostly gone. He also burned time on a FastAPI base64 padding bug. He ultimately moved the step to a hosted model, which held identity much better, and asks whether InstantID or IP-Adapter FaceID could have saved the local approach.
More from Multimodal
- ComfyUI App Mode studio chains FLUX to Qwen to Wan in one tab, demo run costs just $0.52 — no3us · 2026-11-04
- CreatorOS Studio rebuilds open-source Higgsfield on ComfyUI: photo restore, in-UI character LoRA training, 60+ tools, local and free — Key_Cap_8111 · 2026-11-04
- Local Video Pipeline on RTX 3090: 3h15 for 1m22s of 60FPS AI Video — bmallCakeDiver · 2026-11-04
- LayerShop 0.19 Adds Multi-LoRA Stacking and External ComfyUI Model Folders — HonestTicket6013 · 2026-11-04
- Tutorial: Beat-Synced Lyric Videos via Claude Code + Pokee Isaac — Kyrannio · 2026-11-04
- Tencent's WeChat Vision team open-sources WeVisOne-VL, a unified 2B/4B perception VLM — bdsqlsz · 2026-11-04