LoRA Training Center
Train a character once and use it everywhere it fits. One set of pictures (a character sheet or your own photos) becomes a dataset, and from that dataset the hub trains a LoRA for each kind of model you pick. Each one plugs into the parts of the site that run on that model.
Video & scenes Ready
WAN 2.2 · the same LoRA type as @Auntie, @CrissP and @BlondeBree
- @mentions in Scene mode
- Create Video
- Claude connector
Images Beta
FLUX.2 Klein 9B · trained test passed, going into Image Studio
- Image Studio (Klein 9B)
Lip sync Beta
LTX-2.3 · trained test passed (BlondeBree), about 3 h on the GPU
- Lip Sync
- Dance + lip-sync music videos
No training needed
These use your pictures directly:
- Motion Transfer and Swap
- Character Sheets
- Qwen image edits
A LoRA only works with the model family it was trained on, so a WAN LoRA can’t be used in FLUX.2 and the other way round. That’s why the hub trains each family from the same dataset: you set it up once.
Train a character
Builds about 30 training pictures (new angles, outfits and places) from your source, then trains. Each family takes a few hours on the GPU, and other jobs wait their turn while it runs.
Loading your sheets…
Only you see this. Pick one of the hub's influencers: it trains from its card pictures, and when it's done its @handle works for every user (Scene mode, Create Video, Image Studio).
0 / 12 photos (3 minimum)
Your trained characters
Loading…
Advanced: train a style or object LoRA from raw images
New training run
Training takes a while (30 minutes to a few hours) and uses the GPU one job at a time. You can leave the page — it keeps going.