Looks and LoRAs
Training a LoRA
The canvas is the dataset. Curate a board, caption the keepers, pick a trigger word, and train from there.
The canvas is the dataset. You select training images straight off the board, which means the curation you were doing anyway is the preparation, and nobody assembles a zip. ## The two image models | | FLUX.1 | FLUX.2 | |---|---|---| | Images | 4 to 20 | 9 to 50 | | Speed | Faster | Slower | | Quality | Good | Higher | | Extra control | Style-mode toggle | Learning rate exposed | | At generation | Can steer from a reference image | Text only | Start with FLUX.1 while you are learning what your dataset should be. Move to FLUX.2 when you know the set is right and you want the better result. ## Doing it 1. **Curate on the board.** Generate or gather candidates, stack the keepers, delete the rest. This is the step that decides the outcome 2. **Select the images** for training 3. **Caption each one.** There is an auto-describer built into the tool, and its output is a starting point to correct rather than accept 4. **Pick a trigger word.** Five to eight characters, an invented one like `ch41r` 5. **Set the steps**, between 500 and 2000 6. **Train.** It runs 2 to 10 minutes and notifies you when it lands in your library :::tip[The dataset is the whole job] Twelve varied, well-lit, correctly captioned images beat forty inconsistent ones. Vary angle, distance and lighting; keep the subject constant. If half your images are the same three-quarter view, that is the view you will get back forever. ::: ## Captions matter more than people expect A caption tells the model which parts of the picture are the thing and which are incidental. Describe the surroundings and the pose, and use the trigger for the subject itself. If every caption says "a chair on a white background", the model learns that white backgrounds are part of your chair. ## Training a 3D style The same idea, applied to geometry. Pick **10 to 20 textured models** off your canvas and train a Trellis 2 style LoRA on them. Three tiers, and they learn increasingly deep: - **Fast**, texture style only - **Balanced**, adds geometry - **Max**, full style including coarse shape It takes roughly 10 to 20 minutes and needs at least ten usable models after filtering. :::caution[Ask before your first 3D style run] This one is worth a conversation with the Plnty team first. They can share examples of datasets that worked, and a 3D style trained on the wrong ten models is an expensive way to learn what the right ten look like. ::: ## After it trains The LoRA lands in your library and is available to every board in the workspace. Using it is covered in [applying one look to a set](/docs/using-plnty/looks-and-loras/applying-one-look-to-a-set/). :::note[Training is terminal, not a pipeline step] A training run produces no canvas asset. It populates your library and notifies you, which is why nothing appears on the board when it finishes and why you cannot chain it. :::The canvas is the dataset. You select training images straight off the board, which means the curation you were doing anyway is the preparation, and nobody assembles a zip.
The two image models
Section titled “The two image models”| FLUX.1 | FLUX.2 | |
|---|---|---|
| Images | 4 to 20 | 9 to 50 |
| Speed | Faster | Slower |
| Quality | Good | Higher |
| Extra control | Style-mode toggle | Learning rate exposed |
| At generation | Can steer from a reference image | Text only |
Start with FLUX.1 while you are learning what your dataset should be. Move to FLUX.2 when you know the set is right and you want the better result.
Doing it
Section titled “Doing it”- Curate on the board. Generate or gather candidates, stack the keepers, delete the rest. This is the step that decides the outcome
- Select the images for training
- Caption each one. There is an auto-describer built into the tool, and its output is a starting point to correct rather than accept
- Pick a trigger word. Five to eight characters, an invented one like
ch41r - Set the steps, between 500 and 2000
- Train. It runs 2 to 10 minutes and notifies you when it lands in your library
Captions matter more than people expect
Section titled “Captions matter more than people expect”A caption tells the model which parts of the picture are the thing and which are incidental. Describe the surroundings and the pose, and use the trigger for the subject itself.
If every caption says “a chair on a white background”, the model learns that white backgrounds are part of your chair.
Training a 3D style
Section titled “Training a 3D style”The same idea, applied to geometry. Pick 10 to 20 textured models off your canvas and train a Trellis 2 style LoRA on them.
Three tiers, and they learn increasingly deep:
- Fast, texture style only
- Balanced, adds geometry
- Max, full style including coarse shape
It takes roughly 10 to 20 minutes and needs at least ten usable models after filtering.
After it trains
Section titled “After it trains”The LoRA lands in your library and is available to every board in the workspace. Using it is covered in applying one look to a set.