Generate or Edit Image · flux-kontext-pro

Two Kontext nodes ship in this pack - Pro and Max - with identical controls and different model IDs. If you're here, the practical question is "should I have clicked Max?" The answer, most of the time: start here, and only pay for Max when Pro keeps missing.

Kontext itself is the 12B instruction editor that turned "describe the change" into the normal way to edit an image, and it's still the most reliable thing at keeping a face recognisable across passes. It's also infamous for the licence update that came with it, which is a separate argument; the hosted node doesn't care either way.

How it works

The node runs KIE's Kontext endpoint in whichever mode your inputs imply. Prompt only → text-to-image generation. Prompt plus inputImage → the same prompt becomes an edit instruction. Then it polls the job, downloads the result and returns a native ComfyUI IMAGE tensor.

The catch, same as the Max node: inputImage is a string URL, not an IMAGE socket. Kontext's documented page doesn't declare it as media, so the pack leaves it as text and you supply a URL the API can fetch. In practice you get it from the pack's own upload node - wire KIE • Upload Image and take its first_url, or use any public link.

Inputs

prompt - required, English only. For edits, write the change, not the picture: "make the sky overcast and remove the people on the left" beats a full re-description, because the model already has the image.

inputImage - the URL. Leave it as the placeholder and you'll get a generation instead of an edit, which is a confusing way to find out you forgot it.

enableTranslation (default on) - auto-translates a non-English prompt before sending. Turn it off when you're writing English and want your exact wording preserved.

aspectRatio - dropdown, 16:9 by default, six choices. In generation mode it sets the output shape; in editing mode, setting it asks for a re-frame. Leave it out of the edit's way unless you want that.

safetyTolerance - dropdown defaulting to 2. The tooltip distinguishes the modes: generation accepts 0 (strict) through 6 (permissive), editing caps at 2, described as balanced. So a value that feels conservative in generation is the maximum permissiveness for an edit.

outputFormat - jpeg by default. Switch to PNG if the output is going back into the graph for more work; JPEG between passes loses detail you'll want later.

promptUpsampling (default off) - the service rewrites your prompt first. Good for one-liners, bad for prompts you thought about, and it adds latency either way.

uploadCn - picks a China or non-China server for the image upload, based on where you're sitting. Leave the default unless uploads drag.

watermark - pre-filled with a placeholder string, and any non-empty value means a watermark gets added. Clear the field.

Plus timeout_seconds (1200 default) and callback_url (empty unless you run a webhook).

Outputs

image is the edit or generation as a native tensor. Then the pack's standard suit: url for the hosted file, all_urls_json for all returned URLs, task_id for the job, raw_json for the raw response, and credits_consumed / credits_left as floats, mirrored on the node after each run.

Install

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

ComfyUI Manager does the same job if you search "KIE.ai Nodes Next". Restart ComfyUI, then paste your KIE key once in Settings → KIE.ai Nodes Next → Connection - validated against KIE before it replaces a saved key, kept out of your workflow, masked in the UI. On a server you can set KIE_API_KEY and skip the UI entirely; it takes priority.

The install is requests, Pillow, numpy, PyYAML. Nothing to download, because the model is running on someone else's hardware.

Notes from using it

• Kontext drifts. It'll nudge details you never mentioned and get softer over repeated edits. Keep the original in the graph and re-edit from it rather than editing an edit.
• Pro vs Max is a quality tier on the same feature set. Work the prompt on Pro; escalate when you've stopped gaining anything.
• The watermark field is the most common accidental mistake on this node after inputImage, thanks to that pre-filled placeholder.
• First-run "endpoint not resolved" - the pack builds nodes from KIE's live documentation and refreshes the catalog in the background. Restart or hit the catalog refresh node and the endpoint fills in.