FLUX LoRA keeps losing facial identity: is 24 images enough and how to fix it
Minxit08 · reddit · 2026-11-03
A Reddit user troubleshooting a personal-identity FLUX.1-dev LoRA trained via fal.ai: 24 images with varied hairstyles, angles, and scenes, yet generations look like a generic pretty face instead of the actual person — narrower, hollow-cheeked, personality drained.
Open questions: how to balance close-up neutral portraits vs full-body/lifestyle shots; whether a smaller curated dataset beats 24 mixed images; whether facial-proportion captions risk overfitting; how LoRA rank, steps, learning rate, and inference strength affect identity; whether pairing the LoRA with a face-reference adapter beats retraining; and how to objectively evaluate identity consistency across poses and lighting. Useful discussion thread for portrait-LoRA tuning.
More from Multimodal
- ChatGPT Pro conjures a surprisingly charming 'Velvet Kin' mascot for an AI agent — RileyRalmuto · 2026-11-03
- Astrocade launches Avatar Studio v1 for creating and exporting game-ready AI avatars — _amirabs · 2026-11-03
- Solv3R Turns a Looped Transformer Into a Multi-View Geometry Solver With 16× Pipelined Throughput — taiyasaki · 2026-11-03
- OpenAI says 1.5 billion images are generated per week — firstadopter · 2026-11-03
- ComfyUI-LATO2 turns TRELLIS.2/Hunyuan3D meshes into usable low-poly, 990k to 3k faces — sakalond · 2026-11-03
- Comparing Different AI Models' Pixel Art Generation Skills — Angaisb_ · 2026-11-03