Bria FIBO Image Edit

Bria is the commercial-API shop that does the boring image jobs well - background removal, erasing, fill - and Bria FIBO Image Edit is its instruction-editing model with a genuinely unusual twist: it will edit your image from a plain sentence, or from a machine-readable structured JSON prompt, and it hands you the structured prompt it ended up using as a second output. That second output is the reason to notice this node. You can round-trip it: run the edit, read the structured prompt it produced, refine it, and feed it back - programmatic control over a text-editing model, which is rare among the partner nodes.

It's a partner node like the rest: image (and optional mask) go up to Comfy's proxy at api.comfy.org, Bria's servers run FIBO, credits come off your Comfy account, nothing touches your GPU, no model files. The one mechanical difference worth knowing: this node uploads your image to Comfy's storage first rather than base64-inlining it, so you'll see an "Uploading image" step before the edit starts.

Inputs that matter:

• prompt / structured_prompt - at least one must be non-empty, and the node refuses to run otherwise. Plain sentence or JSON, your choice. negative_prompt is there if you want to steer away from something.
• mask - optional, and the tooltip is the whole story: omit it and the edit applies to the entire image. Provide it and you get masked, localized editing - the thing instruction models are usually bad at, which is a real differentiator.
• guidance_scale (3–5, default 3) - higher follows the prompt more tightly.
• steps (20–50, default 50), seed, and moderation - a DynamicCombo; pick "true" and it expands into three toggles (prompt_content_moderation, visual_input_moderation, visual_output_moderation) if you want Bria's content filter on specific legs.

Outputs: the edited IMAGE, and structured_prompt - the exact structured JSON Bria used. Save it to a text file and you've got a reusable, tweakable edit spec.

On moderation, a grounded note: Bria's API (like most of these) runs its own content checks regardless - the toggles control which legs are checked and how strictly, and a flagged request comes back as an error. If you hit an unexplained rejection, that's usually what happened. And since this is a partner node with real per-call pricing, the badge on the node is the current price - read it before you batch fifty edits.

Where it fits in a workflow: the masked-edit capability makes this the pragmatic middle ground in the mask-vs-instruction debate - you get sentence-driven edits and the option to lock the unmasked region down so nothing outside the mask drifts. If you've ever wished Kontext-style editing came with a way to say "but don't touch this part," this is the node that actually has the input for it.