AI Toolkit has a no-code web UI, so you don't have to edit YAML by hand.

New to it? Watch Ostris' 2-minute walkthrough. BFL's own docs: klein LoRA training and a worked example.

• 15–40 images that share one look. Varied subjects/angles, ≥1024 px.
• One .txt caption per image, same filename.
• Caption the content, not the style. Describe what's in the image and nothing about the look. The model picks up the style on its own, so don't write "watercolor / retro / muted palette."
• Start every caption with a made-up trigger word (MYSTYLE7, ZK_TOON) so it can't collide with real words. Use it identically everywhere.

MYSTYLE7. A portrait of an older man with a beard, plain background.
MYSTYLE7. A wide shot of a fishing boat on calm water at dawn.

No time to caption by hand? Auto-caption with any vision model ("describe content only, no style words"), then skim for leaks.

Use the ready config configs/my_lora_klein_4b.yaml (Files tab). Change just three lines marked <<< CHANGE >>>:

• name: your output folder
• trigger_word: must match your captions
• datasets: folder_path: where your images are

Then paste it into the AI Toolkit UI (or python run.py <config>) and hit Start. ~1800 steps takes 30–40 min on a 4090.

Two common mistakes to avoid:

• Keep arch: "flux2_klein_4b" in the config. Without it the trainer crashes (#691).
• Watch the sample images, not the loss. Loss keeps dropping past the point where the images look best, which is usually around step 750–1500. Pick that checkpoint's .safetensors.

Download your .safetensors, open the 🎯 Your LoRA tab, upload it, put your trigger word in the prompt, and compare base ↔ LoRA.

In code, two lines on top of the normal pipeline:

pipe.load_lora_weights("my_lora.safetensors")
img = pipe(prompt="MYSTYLE7. a portrait, studio lighting",
           num_inference_steps=50, guidance_scale=4.0).images[0]

Train your own FLUX.2 [klein] LoRA in ~30 minutes

A LoRA teaches klein a style, character, or look from a handful of images. It's a quick way to give your project its own visual identity, and it earns the 🎯 Well-Tuned merit badge.

This guide is self-contained. You need: ~20 images, a GPU for ~30 minutes (a RunPod RTX 4090 is about $0.50 for a full run), and the config in configs/my_lora_klein_4b.yaml.

Trainer used here: ostris/ai-toolkit — a popular community trainer (one of several; you can use any klein-compatible trainer). Model: FLUX.2-klein-base-4B (Apache 2.0).

Don't want to train? You don't have to. The 🎯 Your LoRA tab in the Space has a dropdown of ready-made klein LoRAs from the Hub, and there are dozens more at huggingface.co/models. Train one when you want a specific look that the ready-made ones don't cover.

Pick your path (easiest first)

AI Toolkit has a no-code web UI — you don't edit YAML by hand unless you want to. Three ways to run it:

RunPod (recommended): deploy the official AI Toolkit by Ostris template on an RTX 4090 (24 GB) or L40S with an ≥80 GB volume. The pod auto-launches the AI Toolkit web UI — you create a job, point it at your images, pick FLUX.2-klein-base-4B, and click Start. The configs/my_lora_klein_4b.yaml in this Space is the same job expressed as YAML, so you can either fill the UI form or paste the config. Ostris has a 2-minute walkthrough video.

Local UI: if you have the GPU, follow the ai-toolkit README — clone, install, npm run build_and_start, open localhost:8675. Same UI as the pod.

Modal: clone ai-toolkit, pip install modal, modal setup, add a READ HF token, then run their Modal training command. Good if you'd rather not rent a pod. Steps are in the ai-toolkit README.

Whichever you pick, the dataset + caption rules below are identical, and you end up with a .safetensors you load in the 🎯 Your LoRA tab.

1. Build a dataset (15–40 images)

A style LoRA is the easy win for a weekend. Collect 15–40 images that share one look — your own art, photos you have rights to, public-domain works (Wikimedia Commons is reliable and license-clean).

• Diverse subjects, angles, compositions. Avoid the same background repeated.
• ≥1024 px on the long edge.
• Name them img (1).png, img (2).png, … each paired with img (1).txt, …

Captions: describe the CONTENT, never the style

This is the one rule people get wrong. For a style LoRA, your captions must describe what is in the image and say nothing about the style — that's exactly what you want the model to infer on its own.

Each caption starts with your trigger word, then a literal description:

MYSTYLE7. A portrait of an older man with a beard, facing left, plain background.
MYSTYLE7. A wide shot of a small fishing boat on calm water at dawn.

Do not write "watercolor", "painterly", "vellum", "retro", "muted palette". If you describe the style in the caption, the model learns to need that word instead of baking the style into the weights.

Don't want to caption by hand? Any vision model (Qwen2.5-VL, GPT-4o, Gemini) can auto-caption with a "describe only content, no style words" prompt — then skim the .txt files and delete any style adjectives that leaked in (they leak ~25% of the time).

Pick a trigger word that is not a real word so it can't collide with the model's vocabulary: MYSTYLE7, RISO_PR1NT, ZK_TOON. Use it identically in every caption and in the config.

2. Get a GPU

RunPod (recommended, ~$0.50/run):

1. Deploy the official AI Toolkit (Ostris) template on an RTX 4090 (24 GB) or L40S. It auto-launches the AI Toolkit UI.
2. Volume ≥ 80 GB mounted at /workspace (weights + checkpoints need room).
3. Upload your dataset folder to /workspace/datasets/my_style/ and the config to /workspace/configs/my_lora_klein_4b.yaml.

Colab: an L4 or A100 runtime works too — pip install ai-toolkit, upload the dataset, point the config at your Drive path.

3. Edit three lines in the config

Open configs/my_lora_klein_4b.yaml and change the lines marked <<< CHANGE >>>:

• name: — your output folder name.
• trigger_word: — your trigger (must match your captions).
• datasets: folder_path: — where you uploaded the images.

Also update the three sample.prompts to use your trigger so the in-training preview images show your style forming.

The one line you must NOT delete: arch: "flux2_klein_4b". Without it ai-toolkit falls back to a Stable Diffusion loader and crashes on a missing-unet error (issue #691). The official BFL example omits it — that's a bug, the included config keeps it.

4. Train

In the AI Toolkit UI: paste the config, click Start. Or from the CLI:

cd /app/ai-toolkit
python run.py /workspace/configs/my_lora_klein_4b.yaml

It checkpoints every 250 steps into /app/ai-toolkit/output/<name>/ and writes sample images alongside. A 1800-step run on a 4090 takes roughly 30–40 minutes.

Watch the samples, not the loss. Loss keeps dropping past the point where the images start to overfit. For most style LoRAs the visual peak is around step 750–1500, not the final step. Open the sample images, pick the checkpoint that looks best, and use that .safetensors.

5. Use it

Download the .safetensors you picked, then in this Space open the 🎯 Your LoRA tab, upload it, put your trigger word in the prompt, and compare base vs your fine-tune at the same seed.

In code it's two lines on top of the normal pipeline:

from diffusers import Flux2KleinPipeline
import torch

pipe = Flux2KleinPipeline.from_pretrained(
    "black-forest-labs/FLUX.2-klein-base-4B", torch_dtype=torch.bfloat16
).to("cuda")
pipe.load_lora_weights("my_lora.safetensors")

img = pipe(
    prompt="MYSTYLE7. a portrait of a person, studio lighting",
    num_inference_steps=50, guidance_scale=4.0,
).images[0]

You trained on base (50 steps). The LoRA also loads on the distilled FLUX.2-klein-4B (4 steps) for fast demos, with mild drift. The Space's LoRA tab uses distilled by default for speed; switch KLEIN_MODEL_ID to base for the cleanest results.

Quick knobs

More detail: BFL klein training docs · training example.