Civitai Community Training Techniques
last edited
Civitai community training techniques come from shared post-mortems and living guides rather than official documentation. Practitioners estimate LoRA quality at roughly 80% dataset and 20% settings. Failed runs are collected in Civitai training post-mortems. The wider method sits in the advanced training reference and the models reference.
Datasets and captions
Community practice puts character datasets at 15–25 images. Style datasets run larger, at 20–40 images. SD 1.5 training starts from 512-pixel squares. SDXL uses 1024-pixel squares.
Captioning matters as much as image quality. For Flux, natural-language captions beat tag lists because they name subjects, attributes, composition, and recurring concepts. Trainers auto-caption, then edit by hand.
Steps and tools
Community experiments converged on 400 steps. Some runs peak at 300 or 500, so interval checkpoints matter. Longer runs can do better at 3k steps than at 4k by avoiding overbaking. These habits follow the LoRA adoption timeline.
Civitai's own default differs. A June 2026 announcement set new trainings at around 2000 steps. That spans 10 epochs. Ostris' AI-Toolkit is favored for sensible defaults; see training tool comparisons, Flux LoRA community techniques, and advanced Flux settings.
On-site training
The Civitai trainer needs no local GPU. Results reach local setups via helper nodes and spread like extensions. Since June 15, pricing is per step rather than per epoch, paid in Buzz. Flux and video bases cost more.
The new studio flow entered beta on September 23, with live progress, epoch generation, and side-by-side comparison. Support for a newer image base arrived on October 17. The classic trainer stops starting new runs on November 16. The cutoff is midnight UTC, and the studio then becomes the default.