comfy

Training a LoRA is the one job in this pack that's a real commitment, and for a lot of people it's also the one they can't do locally - no VRAM for it, or no desire to babysit a multi-hour run. This node moves the whole thing to Civitai's fleet: you hand over training data and hyperparameters, a hosted job trains your LoRA, and you get back a checkpoint per epoch, billed in Buzz. It's the ai-toolkit training pipeline, exposed as a single ComfyUI node.

What it is

comfy (display name) - CivitaiTrainingComfy - from civitai/civitai-comfy-nodes, under Civitai/Training/comfy. The discriminator pins the engine to comfy, meaning the hosted training backend runs the Comfy-style native training pipeline rather than the legacy AI-toolkit path. The tooltips make the pricing model explicit: training is billed per step, each epoch is a delivered checkpoint plus preview samples, and storage is a per-epoch surcharge on top.

The inputs that matter

The schema is enormous, but the ones that decide your run are few:

• training_data_json - required. Your dataset, as JSON. This is where the KB's oldest lesson still rules: dataset curation beats every knob in this node. Caption well, keep the set consistent, and the rest of your settings just don't have to be perfect.
• steps (1–10,000) and epochs (1–200) - the two ways to set training length. The tooltips describe the dance: steps is the primary control and determines pricing; when you supply it, epochs (the number of saved checkpoints) is derived from it, and vice versa. Epochs = downloadable models. Want several checkpoints to pick from? Set epochs; want a fixed budget? Set steps.
• network_dim / network_alpha - rank and alpha. The tooltips are unusually candid: a larger Dim stores more but can learn junk and grows the file; with Alpha 16 and Dim 32 the effective strength is 16/32 = 0.5. Keep them equal to start.
• lr (0.0001), optimizer_type (the tooltip's honest line: "the most commonly used in LoRA learning is adamw8bit"), lr_scheduler - the learning-rate trio.
• trigger_word - the word that activates your LoRA in prompts. Only meaningful for certain ecosystems (sd1, sdxl, flux1, chroma, and friends), per the tooltip.
• train_text_encoder - off by default, and per the KB that's the right instinct for modern models. Turning it on improves prompt understanding but adds time and memory.
• model - CIVITAI_AIR. The checkpoint to train upon; wire a Model Selector. diffusion_model / text_encoder_model / vae_model are the split-model overrides for the native Comfy path.
• continue_from - a previously trained LoRA, to keep training instead of starting from base. Nice for refining.
• batch_size (1–4), resolution (256–1024, default 1024), noise_offset, flip_augmentation, shuffle_tokens / keep_tokens - the augmentation and efficiency controls.

The outputs

• moderation_status - the age-classification / moderation result for your dataset (the node even has requires_age_classification as an input).
• epochs - the list of trained checkpoints. This is your deliverable.
• workflow_id / raw_json - the pack standard.

Installing it

ComfyUI Manager → Civitai Comfy Nodes → Install → restart. Or:

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

Gotchas

• This is the expensive node. Per-step training plus per-epoch storage, all in Buzz. steps is literally described as "the primary control over training length and determines pricing" - treat it as your budget dial.
• Your data leaves your machine - the entire dataset is uploaded to Civitai's fleet. Training is the one operation here where the "cloud" framing is not a convenience, it's the whole point.
• Epochs are checkpoints, not iterations. Save several and pick the best - the last epoch is often overtrained, the KB's longest-running lesson.
• Credentials required - Civitai Auth node, CIVITAI_API_TOKEN, or sidebar OAuth. Training with no balance will fail fast, not politely.
• It's one node, but it's not a beginner's first run. If this is your first LoRA ever, the KB's advice holds: curate the dataset, don't train the text encoder, and don't trust the last epoch.