FLUX.1 Kontext [dev]

FLUX.1 Kontext killed the mask-based editing workflow, or at least demoted it. Kontext is the instruction-editing model: you feed it an image and say "make it raining," "change the jacket to red," "add a mountain behind the building," and it rewrites the image to match - no mask, no inpaint, no denoise dial. By 2026 the community verdict was settled: Kontext (and its siblings Qwen-Image-Edit and FLUX.2 klein) took over most of the job people used masks for, because editing from a sentence beats editing from a hand-painted region for almost everything.

The catch is that "dev" is the open-weights tier, and dev weights are non-commercial with BFL's tightening license. This node sidesteps the whole thing: it's the Kontext dev model running on Runware's cloud, so you get the capability without hosting the weights or shipping a fork of them anywhere.

What you set

Required is positivePrompt - and that's the entire editing instruction. Then wire referenceImages with the image you want changed. From there it behaves like a full image node:

• negativePrompt - has a tooltip warning worth taking seriously: it's ignored when guidance is disabled (CFGScale ≤ 1). Kontext-family guidance-distilled models run low-CFG by default, so your negative may do nothing at all. As the photorealism crowd put it, write constraints as presence ("clean studio background, sharp focus") instead.
• steps - a toggle plus steps_value (flip the switch, then set 1–50). Default "off" means the model picks.
• CFGScale / trueCFGScale - both gated toggles with _value siblings; leave them off unless you know why you're overriding.
• scheduler - the full 76-entry dropdown. Leave it on (default).

The typed sockets are where this node earns its keep: lora, ultralytics (object-detection guidance), acceleratorOptions (caching/distillation speedups), and watermark. These come from the pack's builder nodes under Runware/Params and wire straight in. Output is one image.

The gotcha that defines this model

Kontext edits the whole image. That's the trade: an edit model takes everything into context and emits a whole new image, so the parts you didn't want touched come back close-but-not-identical. Community wisdom is that drift compounds across a chain of edits - each pass nudges the background a bit more. If you need bit-identical unmasked pixels (a logo, a face you must not change), the mask still owns that job; run your instruction edit, then composite the edited region back over the original with a mask.

Install

Standard pack install:

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

Restart, or just install "Runware" via ComfyUI Manager. API key required (Settings → Runware API key or RUNWARE_API_KEY); cost shows on the node title bar. No model files download - everything runs on BFL's weights hosted by Runware, which is the whole point when the local path means either 12GB of weights and a licensing headache or a GPU you don't own.