Generate or Edit Image · flux-kontext-max

Kontext is the model that made instruction editing a commodity - feed it a picture and a sentence, get the same picture with the clothing changed or the background swapped. It's also the model people remember for watermark removal and for how good it was at keeping a character recognisable across edits. This is the hosted Max tier of that, and the node does double duty: no input image means generate, input image means edit.

That's the neat part of the docs page KIE published, and the pack kept it: one node, two jobs.

How it works

You write a prompt. That's required in both modes. Then, if you supply an image, the same prompt is interpreted as an edit instruction instead of a scene description - Kontext consumes references in-context rather than through adapter embeddings, which is why it holds subject identity better than the ControlNet-era approach it replaced.

The gotcha that catches everyone: inputImage on this node is a plain string URL, not an IMAGE socket. Nothing in the pack uploads it for you. You need a URL KIE's servers can reach - either something already public, or one you produced with the pack's upload utilities:

KIE • Upload Image  →  first_url  →  paste into inputImage

The whole node is a thin wrapper over KIE's Kontext endpoint: submit, poll, download, decode to a ComfyUI IMAGE. Expect a straight substitution, not a re-imagining - that's what the model is for.

Inputs worth setting

prompt is required. English only, per the tooltip, and enableTranslation (default on) is what saves you if you type in another language - it auto-translates before the request. If you're already writing English you can turn it off and skip a step.

inputImage - the URL. Empty means text-to-image generation; filled means edit.

aspectRatio is a dropdown defaulting to 16:9 with six options. In generation mode the output follows it. In editing mode, supplying it tells the model to re-frame the result - so if you want the edit in the original's shape, think about this before you leave a default in place.

safetyTolerance is a dropdown defaulting to 2 with seven steps. The tooltip gives the range honestly: 0 (strictest) to 6 (most permissive) for generation, but 0 to 2 for editing, where 2 is described as balanced. So the same number means different things depending on which mode you're in.

outputFormat is jpeg by default, with the other option being PNG. For anything you'll composite or re-edit, PNG is worth the extra bytes - repeated JPEG passes on an image you're editing in stages is how detail dies.

promptUpsampling (off by default) has the service rewrite your prompt before generating. It usually helps thin prompts with nothing to lose and can hurt prompts you carefully wrote; it also costs time.

uploadCn is a server-region switch for the image upload - China or not, chosen by where you are. Leave it unless uploads are slow.

watermark ships with a placeholder value (your-watermark-id), and the tooltip says a watermark is added whenever that field is non-empty. Clear it unless you actually want one.

Plus timeout_seconds and callback_url, the same pair every node in the pack carries.

Outputs

image - native tensor, wire it downstream. url - the hosted result, all_urls_json - every URL returned, task_id - the job ID for follow-ups, raw_json - full provider response, credits_consumed / credits_left - floats reporting cost and balance.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/felipederosilva/ComfyUI-KIE-Nodes-Next

Or search "KIE.ai Nodes Next" in ComfyUI Manager, then restart. Dependencies are requests, Pillow, numpy, PyYAML. Set the key once in Settings → KIE.ai Nodes Next → Connection; it's validated against KIE before being saved over a good key, and it never enters workflow JSON.

Honest notes

Kontext is the previous generation of this idea. It shifts details you didn't ask about and softens an image over repeated passes, which is why the current local favourites for masked editing are Qwen-Image-Edit and Klein - the pack has nodes for those families too. The reason to use Max here is convenience: no 12B model to download, no environment to fight, an edit in twenty seconds from a node you already have. For heavy repeated editing of your own photos, the local path still wins.