FLUX 1.0 [fill] Finetuned

FLUX 1.0 [fill] Finetuned is the inpainting node for people who've trained a model of their own. Take the masked-edit power of the fill node - change only what's under the mask, leave everything else alone - and run it through a model you trained on BFL's servers, so the stuff you're painting in actually looks like your subject instead of a generic approximation.

Same division of labor as every Finetuned node in the pack: training is BFL's job, this node only consumes. You train through the fine-tuning API at api.bfl.ml, get a finetune_id back, and paste it in. There's no training UI here. The classic use case is consistency work - remove an object or extend a background and have the new pixels match the character you've been generating all session, instead of drifting toward whatever the base model thinks a person looks like.

Inputs that matter

• image (required) - what you're editing.
• finetune_id (required) - your trained model's ID; the node errors until this is valid.
• prompt (required) - what should fill the masked region.
• mask (optional) - a MASK input from any mask-producing node. Without it, the node edits the whole frame.
• finetune_strength (0–2, default 1.1) - how hard the finetune dominates the fill. Above 1 pushes hard toward your subject; below 1 lets the base model take over.
• guidance (default 60, range 1.5–100) - fill endpoints like big guidance; this is the same default as the non-finetuned fill node.
• steps (15–50, default 50) - quality vs. cost.
• output_format (default jpeg) - flip to png for lossless.
• safety_tolerance (0–6, default 2).

Output is one IMAGE tensor.

A quirk worth knowing

This node lives in the Flux category in the node menu, not "Flux Finetuned" like its canny and depth siblings. The pack's own code never overrides the category here, so you'll find it under Flux - search for "fill" if you can't spot it. Purely cosmetic, but it trips people up the first time.

Install

Same one-time pack setup: 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 BFL_API_KEY (env var, bfl_api_key.txt beside the pack, or api_key_override). No local models, no heavy dependencies.

Costs stack in two places - training the finetune, then every generation - and BFL's moderation still guards the input. Start with a small cheap fill to confirm the finetune_strength is dialed before you run a big outpainting job; re-generating an entire extended scene because your subject came out weak is exactly how the credit bill sneaks up on you.