Flux Pro Kontext Multi (fal)

Kontext is Black Forest Labs' instruction editor: feed it an image and a sentence, get the same image with the clothing changed, the background swapped, or the watermark gone. Its real strength is character consistency - it consumes reference images in context rather than through adapter embeddings, which made it the backbone of turnaround-sheet and "same character, new scene" workflows. The Multi endpoint is that idea with several inputs at once: give it two-to-four images and one instruction, and it edits across all of them together.

The Pro part matters for why this is a node and not a local model. Flux comes in tiers - Schnell and Dev you can download, but Pro never left the API. So Kontext Pro is only reachable through a bridge like this one: your images go to fal.ai, fal runs the model, you pay per generation, and nothing touches your GPU.

What it actually does

The node ships your images and prompt to fal, Kontext Pro runs remotely, and you get a native IMAGE back. The multi-image conditioning is the feature: combine a character reference with a scene reference and instruct the merge, or keep one subject consistent while restyling around it. Worth knowing the model's history - Kontext is also where the license fight happened, the update that prohibited circumventing content restrictions and turned every NSFW LoRA into a violation. That's why the uncensored crowd moved to Qwen-Image-Edit; if that's your use case, this isn't your node.

The inputs that matter

• prompt (required): the edit instruction, applied across the inputs.
• image_1, image_2 (required): your two primary images.
• image_3, image_4 (optional): additional references when the job needs them.
• guidance_scale (default 3.5): how tightly it follows the instruction versus preserving the originals.
• aspect_ratio: presets from 21:9 to 9:16.
• safety_tolerance: 1–6. Lower is stricter; this is BFL's content filter dial, not a quality setting.

There's num_images (1–4), max_quality, output_format, and a seed. The output is a native IMAGE, so it drops straight into the rest of your graph.

Installing it

ComfyUI Manager → search ComfyUI-fal-API → install → restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/gokayfem/ComfyUI-fal-API.git
cd ComfyUI-fal-API
pip install -r requirements.txt

Restart ComfyUI. No checkpoint to place; you need fal-client and a fal API key in config.ini under [API] (or FAL_KEY in the environment).

Where people get burned

The most common surprise is a content refusal - Kontext's filter is real, and safety_tolerance is how you nudge it, within limits. The pack surfaces fal's actual error text, so a rejection reads as a message, not a broken node. Second, don't over-stuff the inputs: two well-chosen references with a clear instruction beat four fighting each other. And because this is the Pro tier, it's priced accordingly - iterate with a fixed seed so the cache saves you from re-paying for identical calls, and only fan out num_images once you like the direction.