FLUX.1 Kontext [dev] (custom model)

FLUX.1 Kontext [dev] is the model that made instruction-based editing feel like a party trick. Feed it an image and a sentence - "change the shirt to red," "remove the watermark" - and it edits in place, keeping the character's identity instead of redrawing a stranger. The community found it genuinely unsettling at watermark removal, and it became the go-to for turnaround-sheet character consistency. The catch has always been the footprint: Kontext is a full FLUX-scale model that wants a serious GPU.

This node deletes that problem. RunwareArch_flux_1_kontext_dev is the pack's "custom model" wrapper for the Kontext architecture - it defaults to the dev checkpoint (runware:106@1) but you can point the model field at any Kontext-architecture checkpoint sitting in Runware's catalog. The actual denoising happens on their servers; you get back a normal IMAGE.

How it works

Kontext is architecturally interesting: it consumes reference images in-context rather than through adapter embeddings like IP-Adapter. The reference image and the instruction prompt both ride in as conditioning, which is why identity sticks so well. Mechanically, this node is just a taskType: imageInference request over REST - the SDK sends your prompt, image, and settings, Runware runs the model, and the node downloads the result into an IMAGE tensor.

Inputs that matter

• model - the AIR of the checkpoint. Defaults to the Kontext dev model, so most people never touch it.
• positivePrompt - the instruction, written as what you want the image to become.
• referenceImages - the source image(s) to edit, wired in from a Load Image node. For multiple inputs, batch them or chain Load Image nodes; Kontext accepts a set.
• negativePrompt - honored only when CFG is above 1, so don't bother writing one if you run with guidance disabled.
• steps / scheduler - steps is an off-by-default toggle that reveals steps_value; leave it off to use the model default. The scheduler dropdown has ~76 options, which is honestly more than you need.
• lora - a typed socket for the pack's LoRA builder, if the checkpoint you picked is a fine-tune that expects one.

Install

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

Or ComfyUI Manager → search Runware → Install → restart. Then add an API key in ComfyUI Settings → Runware API key (or RUNWARE_API_KEY). Prepaid, top-up style; the title bar shows each run's cost and NSFW check result.

Gotchas

• No key → instant error. Check the console for "No Runware API key" before blaming the node.
• The dev-tier Kontext license came with a clause about not circumventing content restrictions, which torched a lot of the community's LoRA ecosystem. Fine for legit editing, just know what you're standing on.
• The image output does nothing until it's wired into Preview Image or Save Image - the classic "ran but saw nothing" failure.

If you want Kontext's editing without babysitting a 12B model, this is the cleanest way to get it. You trade a per-run cost for never thinking about VRAM again.