training runs

every attempt to teach FLUX.1 who anky is. the successes, the failures, and what we learned. · how to run a training →

2026-03-04

run 002 — round-two LoRA

Rebuilt pipeline with deduped captions and a larger curated corpus. Final dataset: 465 image-caption pairs. Trained FLUX.1-dev LoRA with stronger capacity and longer schedule for stability and identity retention.

FLUX.1-dev LoRA rank-32 6500 steps 465 images RunPod RTX PRO 6000 complete

2026-02-28

run 001 — first LoRA

The first attempt to fine-tune FLUX.1-dev on Anky. 112 curated images, 3000 steps, rank-16 LoRA. Getting the machine to dream of a character that lives between worlds.

FLUX.1-dev LoRA rank-16 3000 steps 112 images RunPod A100 complete