Civitai Training Post-Mortems

last edited

Civitai training post-mortems are a distinct genre: creators publish failed runs with full settings, so others can see which change fixed what. Success-only tutorials omit that evidence. Each Civitai post-mortem pairs a failure with its fix. The settings consensus lives in Civitai community training techniques, and the wider method in the advanced training reference.

The v3.0 overfitting failure

The v3.0 style LoRA used 56 curated images. It copied them pixel-for-pixel instead of learning a look like the five style families. The images were 2048px, but resolution was not the cause. The settings were tuned for 1024px images. A UNet learning rate of 5e-4 then "flash fried" information-dense images.

The Civitai fix starts with a lower rate of 2e-4. A rate of 1e-4 is the cautious choice. Use cosine or linear schedulers instead of cosine_with_restarts. Long runs invite memorization. Twenty or more cycles is too many, so the suggested range is 8–12.

Memorization and diversity

The two-week Flux post-mortem on Civitai found image-specific memorization whenever repetition was high or concept coverage narrow. More variations of the target concept improved generalization; see advanced Flux settings.

Repeated characters can leak into a style LoRA when captions omit their identities, so one face, outfit, or pose appears across unrelated prompts. Common causes of failure in these Civitai reports are small datasets with too few steps, excessive intensity, repeated subjects, weak captions, and keeping only the final checkpoint.

Lessons

Save a checkpoint every 1–2 epochs. The best usable version often lands between epochs 4 and 8, before memorization sets in. Version-control 3 assets: the config file, the output folder, and the dataset, so changed runs stay comparable. Script changes appear in Kohya SS updates, learning paths in ComfyUI learning resources, and the backlash against low-effort models in AI slop sentiment.

Sources

Both reports were published as creator articles on Civitai with full settings, sample images, and the changes made between runs. Civitai articles: the real reason a v3.0 LoRA failed and the Flux LoRA settings and dataset post-mortem, which records two weeks of tests.