Free Site ImageLoRA.dev Tracks LoRA Trainer Support, Character LoRA Settings, and GPU Rental Prices

SirToki · reddit · 2026-10-20

A Reddit user launched a free site, imagelora.dev, continuously tracking which LoRA trainer (ai-toolkit, OneTrainer, musubi-tuner, kohya sd-scripts, SimpleTuner) supports which model, recommended settings for character LoRAs, and GPU rental prices. The motivation: after FLUX 3, MOSS, and KREA2, the answer to 'which trainer/setting for X' changes constantly.

The site offers a model×trainer matrix (cells with no source marked 'unknown'), VRAM notes where trainers publish them, and live GPU rental prices fetched from RunPod's and Vast.ai's public APIs (with fetch dates and the author's referral links). The author also shares his own Krea 2 recipe via ai-toolkit: LoKr, automagic3, lr 1e-4, best checkpoint around step 5,500—full YAML on the site.

Much of the site is maintained by AI automations and a persistent agent, with the author manually adding findings from Discords and his own runs. It's a v1 with many 'unknown' cells; contributions with sources are welcome.

Original post →

More from Multimodal

Multimodal channel →