flux-dev-fast

This is the training node for people who don't want to think about training. Where the Flux2Dev nodes hand you steps, learning rate, and captions to tune, the flux-dev-fast recipe hides almost all of it. There's no steps field, no learning_rate, no default_caption - you give it a base model, a dataset URL, a count, and a name, and the "fast" recipe handles the rest on Civitai's servers. It's the training-node equivalent of a point-and-shoot.

That makes it a genuinely different animal from its siblings, and the right one for a specific person: someone training their first LoRA, or someone who wants a quick concept test ("will this style even take?") without learning what min-SNR gamma is. The trade is that you can't fix a bad result by adjusting hyperparameters - because you can't see them.

How it works

Same hosted-training machinery as the rest of the pack: model (required, CIVITAI_AIR from a Civitai Model Selector - the "primary model to train upon"), training_data (required, a URL), and training_data_images_count (required, the cost input). Then just:

• lora_name - default my-lora.
• sample_prompts_json / negative_prompt - the sample images the trainer generates so you can judge the result.

That's the entire knob set. No steps, no LR, no captions. The recipe is tuned by Civitai, and "fast" means both "fewer decisions for you" and presumably "fewer steps per run" - which is exactly why this is the cheapest-feeling training node in the pack, in both senses.

Outputs: moderation_status, epochs, sample_images_prompts, sample_input_images, stored_as_assets, eta, workflow_id, raw_json.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/civitai/civitai-comfy-nodes.git
pip install -r civitai-comfy-nodes/requirements.txt

Or ComfyUI Manager → "Civitai Comfy Nodes" → Install → restart. Authenticate; training bills in Buzz.

Gotchas

Know what you're giving up. The KB's LoRA-training doc is blunt that the quality levers live in the dataset, not just the hyperparameters - and this node removes the hyperparameter half of the equation. If your result comes out weak, there's no steps or learning_rate to fix it; your only knobs are the dataset itself (more images, better captions, more consistent subject) and the sample prompts. If you need fine control, the Flux2Dev node - or the kohya node in this pack - is the one that gives it to you. This node is best understood as the onboarding rung: train one, see how it feels, then graduate to the node with the knobs. The dataset-URL requirement also means your training data has to be reachable by Civitai - a local folder won't do.