No dataset folder loaded yet
Click File → Open dataset folder above to get started.
v2.0.0
Osmium Workshop Dataset Manager
Sort:
Editing Stats
Curate
Trainflow
Train a LoRA for Anima or Janima
Nothing starts until you press Start Trainflow. Training runs in the background and keeps going if you close Osmium; reopen it and this tab reconnects to the progress.
Continue training
Not happy with where a run ended? Pick one of this project's checkpoints and Start Trainflow trains on from there instead of starting over.
Log
Not started.
Previews
Previews appear here as training samples them.
Checkpoints
None yet.
Anima-TrainFlow folder
Your Anima-TrainFlow folder (the one with python_embeded and training). Osmium uses its Python and trainer; its own app doesn't need to be open. Tagging and bucketing your dataset use Osmium's own WD14 and Bucket Images.
Not set
Dataset
Loaded dataset
No dataset loaded
Extra image folders
Models (set once)
Pick your own copy of Anima or Janima and its text encoder and VAE.
DiT
Qwen3 text encoder
VAE
Training
Advanced (power users)
LoRA shape
Precision
Schedule & flow
Regularisation
Performance & VRAM
Preview sampling
Preview prompt
SynthDat Overseer
Generation
No image generated yet.
Accept adds this image + its tags to the dataset. Reject still saves it, but straight into the dataset's Disabled/ folder — out of the way, but recoverable later like any other disabled image.
Tags this will save:
ComfyUI connection
Run on
ComfyUI folder
Not set
Attention
Reference image
Pick an image showing the pose you want — WD14 interrogates it so you can copy just the pose tags into the Pose field below via ControlNet (Anima-LLLite), with the character LoRA guiding what actually gets drawn.
Pick an image showing the pose you want — WD14 interrogates it so you can copy just the pose tags into the Pose field below via ControlNet (Anima-LLLite), with the character LoRA guiding what actually gets drawn.
No reference image picked yet.
Model & LoRA setup
LoRA stack (the character LoRA and anything else you want applied):
ControlNet (Anima-LLLite)
Applied against the reference image whenever one is in use — this is what strong-arms the LoRA's character into the reference pose. Skipped entirely if "I don't want to use a reference image" above is checked.
Applied against the reference image whenever one is in use — this is what strong-arms the LoRA's character into the reference pose. Skipped entirely if "I don't want to use a reference image" above is checked.
Reference image resize
This is what ControlNet actually conditions on — the padded/resized preview above shows exactly this.
This is what ControlNet actually conditions on — the padded/resized preview above shows exactly this.
Prompt fields
Character details: