Skip to main content
Reference images work well for photographic qualities and broad aesthetics. They average out when you try to capture a really specific subject or visual language. When that happens, train a LoRA. Time: 8 minutes (plus 20-40 minutes for training to complete) You’ll need: 10-30 reference images with a consistent style or subject, and a Hobby or Pro subscription
For the underlying concepts, see Custom LoRA models.

Train a LoRA on the dashboard

LoRAs are trained from the dashboard rather than from a workflow.

​
When to reach for a LoRA

A LoRA is a small adapter trained on your images. Once trained, every generation is biased toward that look without you having to prompt for it. Use a LoRA when reference images alone aren’t enough:
  • An artist’s style: Hansmeyer’s subdivision work, a specific architect’s tectonic language
  • Your studio’s aesthetic: accumulated visual identity across past projects
  • A repeatable transformation: turning raw renders into polished photographs — this uses an edit trainer trained on before/after pairs instead of a flat image set
For comparison: ControlNets enforce structure (depth maps, edges); IP-Adapters condition on a reference image without training. Different problems. For aesthetic capture, LoRA is the right tool.

​
1. Collect 10-30 training images

The images should all share either:
  • A consistent style (illustration, photographic look, line drawing)
  • Or a consistent subject (a specific person, a specific object)
Don’t mix. The LoRA learns the common element across all images, so consistency is what makes it work — a smaller set with one coherent look beats a bigger set with variety in it. Use one image per scene or project so the LoRA learns the look rather than memorizing a favourite subject. Crop to the part you want captured. Strip out distracting backgrounds where you can.

​
2. Start a training run

There are three ways: From the canvas (fastest if you’ve already gathered images in a workflow):
  1. Paste or drop your training images onto the canvas
  2. Select all of them
  3. Open the menu → Training
  4. Click Train LoRA: the dashboard opens with the images pre-loaded
From the dashboard:
  1. Visit runchat.com/dashboard/training
  2. Click Train LoRA
  3. Upload images directly
With an agent (MCP): an agent connected to the Runchat MCP can run the whole loop — prepare the dataset (including synthesizing before/after pairs for edit training with a canvas workflow), start the run with start_training, monitor it with list_training_runs, and test the result on the canvas. Ask it to load the edit-lora-training skill.

​
3. Configure the training

In the training dialog:
  • Training endpoint: Flux Klein 9B (5,000 credits) is the cheap, fast option — ideal for a first run or iterating on a dataset. Flux 2 (16,000 credits) trains on FLUX.2 [dev] for maximum quality. The Edit variants of both take before/after pairs instead of a flat set.
  • Trigger word: a phrase you’ll add to prompts to activate the LoRA. Use an invented token plus “style” — e.g. MYL0RA style — something distinctive that isn’t a real word.
  • Description: a label so you can find this LoRA later
Click Start Training. The dialog shows the exact credit cost before you confirm; a run takes 20-40 minutes.

​
4. Apply the LoRA in a workflow

Once training finishes:
  1. Refresh your browser to load the new LoRA into your profile
  2. In a workflow, add a Create node and switch the model to a LoRA-capable model matching your trainer: FLUX 2 Lora Edit for Flux 2 trainings, or a Klein 9B LoRA model for Klein trainings
  3. In the node settings bar, click the LoRA dropdown (defaults to “No LoRA”)
  4. Pick your trained LoRA from the list
  5. Include your trigger word somewhere in the prompt — prompts that reuse your training-caption phrasing work best
  6. Run
The output should reflect your training style.

​
5. Tune the strength

If the result is too subtle (LoRA influence too weak) or overpowering (everything pushed into the trained look), adjust the LoRA scale in the settings bar:
  • 0.6–0.8 usually applies the style selectively while keeping neutral colors neutral
  • 1.0 and above pushes the whole image into the trained look
Guidance and inference steps are the other parameters worth experimenting with. Duplicate your node with alt-drag and test values side by side, comparing against your reference images rather than judging outputs in isolation.

​
Why this matters

Once you have a LoRA for your studio’s aesthetic, every render you produce can be biased toward that look without writing increasingly long reference-heavy prompts. It’s the difference between describing a style every time and showing the model enough examples that it learns the style itself.

​
Next steps