Flux LoRA Trainer (fal)

Most nodes in this pack use a model. This one trains one. FluxLoraTrainer_fal takes a handful of your images, runs a Flux LoRA training job on fal's GPUs, and hands back a lora_file_url - a link to the finished .safetensors. No local training rig, no CUDA wrangling, no renting a pod. You feed images in one end and get a usable Flux LoRA out the other.

This is the front half of a loop that lives entirely on fal: train here, then paste the resulting URL into FluxLora_fal's lora_path and generate with your new adapter. The whole thing never touches your GPU.

What actually decides quality

Here's the part beginners skip and regret: the dataset and captions outweigh every knob in the trainer. That's been true across every architecture change since 2022. A tight set of 15–25 good, varied images beats a big sloppy one. Flux Dev has one useful quirk here - it tolerates a slightly dirty dataset for a single-subject LoRA better than a base model would, though it still falls apart if you try to cram three subjects into one adapter. Train one thing at a time.

The inputs that matter

• images - your training set, wired in as an IMAGE batch. This is the single biggest lever on the result.
• trigger_word - the token you'll type later to summon the concept. Pick something distinctive that isn't already a common word.
• steps - defaults to 1000. A reasonable starting point; more steps risk overfitting (the LoRA memorizes your photos instead of generalizing).
• is_style - flip this on when you're training a style rather than a subject; it changes how the job treats the material.
• create_masks - on by default, it auto-masks the subject so the model learns your subject and not its backgrounds.

There's also images_zip_url if you'd rather pass a pre-built dataset ZIP by URL, plus preprocessing flags for already-prepared data. The single output is lora_file_url - a string, not an image.

Installing it

ComfyUI Manager → search ComfyUI-fal-API → install → restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/gokayfem/ComfyUI-fal-API.git
pip install -r ComfyUI-fal-API/requirements.txt

Nothing to download locally. Set your fal key (config.ini, or export FAL_KEY=your_key; keys at fal.ai/dashboard/keys), restart, and find the node under FAL/Training.

Where people get burned

Training is the most expensive thing in this pack - a full run bills real fal GPU time, not a few cents like an image. So get the dataset right before you spend: 1000 steps on a bad set just gives you a confident bad LoRA. Curate first.

The second trap is watching for the output. The node returns a URL, not a file that magically appears in your models folder - copy that lora_file_url, and either download it or feed it straight into FluxLora_fal. Lose the URL and you've paid for a LoRA you can't find (the pack's provenance and job-inbox utilities can help you recover a past request). And keep the license in mind: LoRAs trained on Flux Dev inherit Dev's non-commercial terms, so a LoRA you trained here isn't automatically clear for commercial redistribution even though the images you generate with it are yours.