cinemasetfreeSources, weights and combining
LoRAs come from Replicate, Hugging Face and CivitAI. Weight controls how strongly one applies; start around 0.7–0.8. Stacking works, but the weights have to come down as you add more.
The three sources
| Source | What it’s good for |
|---|---|
| Replicate | Hosted and versioned, so a LoRA that worked last month still works. The most reliable option. |
| Hugging Face | The largest open collection. Quality varies enormously, so read the model card and check the base model matches. |
| CivitAI | Strongest for style and character LoRAs. Check the licence before using one commercially. |
Weight
Weight is how strongly the LoRA applies. 0.7–0.8 is a good starting point for most: enough to take effect, not enough to overwhelm the base model.
Push past 1.0 and the LoRA starts dominating: colours flatten, compositions repeat, everything drifts toward the training set. Below 0.4 it’s usually not doing anything you’d notice. If a LoRA isn’t working at 0.8, the answer is usually a different LoRA rather than a higher number.
Combining
Several LoRAs can be attached at once, a character and a style say. Their effects add, so weights that were right individually will be too strong together. Bring each down; two at 0.5–0.6 often beats two at 0.8.
Combining two LoRAs that fight over the same thing, two style LoRAs or two faces, produces mush. Combine things that occupy different territory.