Introduction

fal just opened LoRA training for FLUX 3 Image Klein, Black Forest Labs' compact image model. As of October 26 (UTC) you can upload a zip of images to teach Klein a subject, a product, or a style, or upload before/after pairs to teach it an edit you keep repeating. The trainers and the matching LoRA inference endpoints are live on fal's API right now, and the trained weights come back to you as a safetensors file.

What shipped

fal launched four endpoints together:

  • Klein LoRA Trainer: text-to-image training from a .zip of images, with optional per-image captions.
  • Klein Edit LoRA Trainer: trains an editing LoRA from before/after pairs, with up to 10 input images per pair.
  • Klein Base Text to Image LoRA: generate with up to 3 LoRAs stacked in one request.
  • Klein Base Edit LoRA: edit or compose from up to 10 reference images, again with up to 3 LoRAs.

fal says your LoRAs work on both Klein and Klein Base, and that Klein's bounding-box layouts still work while a LoRA is loaded. The model pages recommend the Base (undistilled, CFG) endpoints for inference.

The numbers

Here's what the trainer pages list:

  • Steps: 100 to 10,000, default 1,000.
  • Learning rate: default 1e-4.
  • Training resolution: 512, 768, or 1024 (default 1024, aspect ratio kept).
  • Text-to-image training price: $0.0043 per step, so the default 1,000 steps cost $4.30 (100-step minimum).
  • Edit training price: $0.002 per step multiplied by how many references sit in each pair, from x2.11 with one input up to x23.44 with ten. That works out to $4.22 for 1,000 steps with one reference, or $13.90 with four.
  • Inference with LoRAs: $0.035 per megapixel of output, so a 1024×1024 image is $0.035. LoRA scale runs from 0 to 4.

How to prep your dataset

Subject or style LoRA

Zip your images. Any image named ROOT.png (or jpg, jpeg, webp) can come with a caption file ROOT.txt. Images with no caption use your trigger_phrase, so set one if you skip any captions.

Edit LoRA

Name each pair ROOT_start.png for the input and ROOT_end.png for the edited result. You can add extra inputs as ROOT_start2.png, ROOT_start3.png, and so on, up to 10, plus an optional ROOT.txt with the edit instruction. Pairs can have different numbers of inputs.

A tidy edit dataset looks like this:

mug01_start.jpg
mug01_end.jpg
mug01.txt
mug02_start.jpg
mug02_end.jpg
mug02.txt

Kick off a training run

Host your zip somewhere fal can fetch it, then call the trainer:

curl --request POST \
  --url https://fal.run/blackforestlabs/flux-3/klein/base/trainer \
  --header "Authorization: Key $FAL_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "images_data_url": "https://example.com/my-style-set.zip",
    "trigger_phrase": "in the style of mystyle",
    "steps": 1000,
    "resolution": 1024
  }'

You get back lora_file (the safetensors weights) and config_file. Pass the LoRA path in the loras list on the Klein Base text-to-image or edit LoRA endpoint, with a scale.

Why it matters

The edit trainer is the interesting one. If you shoot product photos and always want the same treatment, like a clean backdrop swap, a lighting change, or a packaging restyle, a few before/after pairs can now teach Klein that move. Being able to stack three LoRAs with bounding-box layouts means one brand-style LoRA, one product LoRA, and an exact placement can all go into a single request.

My tip: start small. Train a short run on a handful of pairs at 768, check the results, and scale up steps and references only once the edit holds. Edit training costs more with every extra reference per pair, so a lean test keeps your budget friendly.

Original Source

Conclusion

Klein just picked up its own custom-training path: subject and style LoRAs, edit LoRAs built from before/after pairs, and up to three stacked LoRAs per call, all on fal for a few dollars per run. If there's a look or a retouch you keep doing by hand, this is a good week to teach it to Klein instead.

Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=fal-flux-3-klein-lora-trainer-text-to-image-and-edit-loras

—Aurelia ♡