FLUX 1.0 [canny] Finetuned

FLUX 1.0 [canny] Finetuned is the regular canny node plus one crucial extra: it runs against a model you trained on BFL's servers. Same edge-controlled composition, same "keep the structure, re-skin the content" idea - but now the content is skewed toward your finetune, so you can put a specific face, character, or product into whatever composition the canny edges dictate.

The thing to understand before you use it: you don't train here. The ComfyUI node has no training UI and no dataset loader. You train a model through BFL's separate fine-tuning API (their web console / API at api.bfl.ml, the same place you manage your key), it spits out a finetune_id, and you paste that ID into this node. That's the whole divide: training is BFL's business, and this node is the consumer side.

Inputs that matter

• prompt (required) - describes the content; the finetune pulls it toward your trained subject.
• control_image (required) - the reference whose edges constrain the composition.
• finetune_id (required) - the ID of a model you already trained. Nothing happens until this is valid.
• finetune_strength (0–2, default 1.1) - how hard the finetune influences the result. Above 1 amps up your subject, below 1 lets the base model drift back in. If your subject barely shows up, this is the first knob to raise.
• guidance (default 30) and steps (15–50, default 50) - same role as the plain canny node.
• output_format (default jpeg) - the finetuned endpoints default to JPEG to keep the response small; flip to png if you want lossless.

Output is a single IMAGE tensor, ready for Save Image or any downstream node.

One quirk worth knowing: this variant drops the low_threshold / high_threshold dials the plain canny node has - the server uses its defaults, so edge sensitivity isn't adjustable here. If you find the edges too dense or too sparse, that's a reason to reach for the non-finetuned node and composite instead.

Install and key

Same pack, same steps, once: ComfyUI Manager → Black Forest Labs API Nodes, or git clone https://github.com/black-forest-labs/bfl-comfy-nodes into custom_nodes, restart, and set your key via BFL_API_KEY, a bfl_api_key.txt beside the pack, or api_key_override. No models to download - the finetune and the base model both live on BFL's side.

Where people get burned: pasting a typo'd or stale finetune_id (the API rejects it with a message), and forgetting that finetuned generations cost the same per-image credits as everything else - plus the training itself bills separately. Moderation also still sits in front, so safety_tolerance (0–6, default 2) is your only lever there. And if your finetune barely shows through, check finetune_strength before you retrain anything; it defaults to 1.1, and some subjects need a stronger nudge.