Skip to main content
Training custom models requires a Hobby or Pro subscription

When to reach for a LoRA

Reference images work well for photographic qualities and broad aesthetics. They average out when you try to capture a really specific subject or a distinctive visual language: an artist’s work, your own illustration style, a unique architectural vocabulary. A LoRA is a small adapter trained on your images. Once trained, every generation is biased toward that look without having to prompt for it. It is the difference between describing a style and showing the model enough examples that it learns the style itself. Runchat trains two kinds of LoRA:
  • Style / subject LoRAs learn a look (or a specific subject) from a flat set of images. What the LoRA learns is decided by your images and captions: consistent style across varied subjects teaches the style; one subject across varied contexts teaches the subject.
  • Edit LoRAs learn a transformation from before/after image pairs — for example “flat CGI render in → professional photograph out”. Because the pairs also demonstrate what must not change, edit LoRAs preserve geometry and content far better than prompting alone. Reach for one whenever the task is “improve/convert X while keeping it X”.
For comparison: ControlNets enforce structure (depth maps, edges); IP-Adapters condition on a reference image without training. Different problems, often stacked. For capturing an aesthetic or a repeatable transformation, a LoRA is the right tool.

Available trainers

The Flux 2 trainers produce LoRAs for FLUX.2 [dev] — the strongest realism, especially on entourage like trees and people. The Klein trainers produce LoRAs for FLUX.2 [klein] 9B — much cheaper to train and to run (the distilled Klein models generate in 4 steps), which makes them ideal for iterating on a dataset before committing to a Flux 2 run. The exact cost is always shown in the training dialog before you confirm.

Training a LoRA

  1. Generate or upload the images you want to train on, then open Gallery from the dashboard sidebar.
  2. Hover over each image you want to include and click the checkbox in the corner to select it. Pick 10 to 30 images that share one consistent look or one consistent subject. Consistency is what makes a LoRA work, so don’t mix looks in a single training run.
  3. With images selected, click Train LoRA in the selection bar.
  4. Configure the training options in the dialog and click Start Training.
You can also start from the Training page: click New model > Train from gallery. For edit trainers, images are uploaded as pairs using the filename convention 0001_start / 0001_end: the start image is the input (e.g. a raw render) and the end image is the target output (e.g. the finished photograph). The dialog names files for you in this pattern and validates the pairs before submitting.
The Train LoRA Model dialog with the training endpoint, name, selected images and trigger word

Training configuration

  • Select Training Endpoint: which trainer to use (see the table above). Edit trainers expect paired images; style trainers take a flat set.
  • Trigger Word: a phrase that activates the LoRA in your prompts. The convention is an invented token plus “style” — e.g. GLDNHR style — so the LoRA doesn’t collide with words the model already knows. For edit trainers the trigger word also serves as the default instruction for pairs without their own caption.
  • Name: a label so you can find this LoRA later on the Training page

Training with an agent (MCP)

The entire training loop can be run by an agent connected to the Runchat MCP — this is the most powerful way to build edit LoRAs, because the agent can synthesize the before/after pairs for you with a canvas workflow:
  • list_training_runs lists your runs and the available trainers with costs
  • start_training starts a run from a hosted dataset zip
  • The edit-lora-training skill teaches the agent the full recipe: building and QA-ing a paired dataset in a workflow, captioning it, starting the run, and testing the trained LoRA on the canvas with scale and prompt sweeps
Ask your agent to “load the edit-lora-training skill” to get started.

Checking on a training run

Click Start Training to zip the images and submit the job. Track its status on the Training page (or via list_training_runs from an agent). Filter by Status (Ready, Training, Failed) or Base model. Runs typically take 20–40 minutes depending on the trainer. Each model’s … menu has Details and logs, Copy trigger word, Copy model URL (once it’s ready) and Delete….

Using a trained LoRA

Once training completes:
  1. On the Training page, open the model’s … menu and choose Copy model URL.
  2. In a workflow, install a LoRA-capable model from Fal that matches the trainer: open the node picker, click Add Fal, and paste the model’s Fal URL (for example a Flux 2 LoRA endpoint for Flux 2 trainings, or a Klein 9B LoRA endpoint for Klein trainings). Add a node with that model.
  3. Ask the agent to apply your LoRA to the node, for example: “Apply this LoRA to the image node at scale 0.8: <model URL>”. The agent sets it in the model’s LoRA settings.
  4. Include your trigger word in the prompt. For best results, prompt with the same phrasing you used in the training captions.
  5. Run the node, then tune the LoRA scale: 0.6–0.8 usually applies the style selectively; higher values push the whole image into the trained look.

Installing a custom LoRA

To use a LoRA trained somewhere else, go to the Training page and click New model > Add from URL. In the Install custom LoRA dialog:
  • Paste a Hugging Face or Civitai link into Auto-fill from URL (optional) and click Parse to fill in the details, or fill them in yourself
  • Model URL: a direct URL to the LoRA file (.safetensors)
  • Name and Trigger word
  • Base model: the model family the LoRA was trained for
  • Example image URLs (optional): preview images, one URL per line
Click Install LoRA. The LoRA appears on the Training page alongside your own models, and you use it the same way.

Next steps

  • Train a custom LoRA: a worked example with reference images and tuning tips
  • Models: how to install and use generative models
  • Credits: full cost breakdown across model types