FLUX.2 Klein 9B · Edit

There's a moment in the Klein story that explains this node better than any spec sheet: someone asked whether Kontext was still relevant, and the answer was "I would consider Kontext obsolete at this point. Both Qwen Edit and Klein are much better." Klein 9B Edit is what made a whole category of dedicated editing models look like legacy software. Here it costs about 2.44 credits ($0.0244) a shot, on someone else's GPU.

What it does differently

Most people meet image editing as inpainting: paint a mask, regenerate inside it. Klein 9B Edit is instructed instead of masked - you hand it one to three source images, describe the outcome, and it works out what to change. Give it one photo and ask to remove the trash can; give it two and ask to put the person from one into the room from the other. That multi-reference merging is the capability that kept Klein in workflows for most of 2026, and it's the part no local 6B rival could match.

image_urls is an Autogrow socket: one image input, and each connection spawns another, up to three. Order matters - the service labels them image 1, image 2, image 3 and you can address them by number in the prompt, so be explicit: Use the face from image 1 and the outfit from image 3; keep everything else from image 1.

Prompting an edit properly

This is where hosted nodes can't save you. Klein's encoder is a Qwen3 LLM and reads your prompt as an instruction - but the failure mode people hit is writing a scene description and getting a new scene back instead of an edit. The fix that circulated widely is to prompt the file, not the picture: treat it like a list of photographic commands. Remove the blur and noise, correct the exposure, lift the shadows, add micro-contrast gets a cleaned-up version of your photo. A woman standing in a sunlit kitchen gets you a stranger in a kitchen. When you do want a scene change, name what must survive: Zoom out until the shoes are visible. Maintain the face. Maintain the posture.

Weighting syntax does nothing here. Neither does masterpiece.

The rest of the fields

aspect_ratio defaults to auto, which preserves the shape of the first source image - correct for almost every edit. Choosing a ratio explicitly reframes the result, which is a feature when you mean it. resolution is 1k or 2k and, on the edit endpoints, the rate is flat rather than tiered, so 2k currently costs the same as 1k; the badge is the live truth. seed starts fixed to stop duplicate billing, and control after generate → randomize is the switch you flip to try again.

Outputs

images - a standard IMAGE batch, ready for Save Image or a colour-match pass. task id - string; the handle for the API console and for support if something looks wrong.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/popnow-ai/popnow-comfyui
# restart ComfyUI

Manager → PopNow → Install does the same thing with less typing. Then a key from popnow.ai/api/console/keys into Settings > PopNow, and a second restart so the node list is built from your account's catalog. There is nothing to pip install; the pack declares no dependencies. It does require a ComfyUI recent enough for the V3 node API - the whole pack is comfy_api.latest nodes, including the Autogrow inputs.

Honest weaknesses

Identity drift. Klein 9B preserves a face less faithfully than Qwen-Image-Edit. It looks more real, and it changes the person more. If likeness is the deliverable, chain small edits or use a consistency LoRA rather than one big instruction.

Warm shift. Results often come back a touch warmer than the source. Match it back with a histogram/colour pass; you cannot negative-prompt it away on this endpoint.

Batching image_urls. Each socket can carry a whole batch and every image in that batch is sent as its own reference. A 4-image batch plus one more socket is five references into a model that takes three, and the API will reject the run with a 400 naming the field.

Everything else about this node is the boring kind of good: it's the editing default, and at 2.4 cents a try you can afford to be picky about the result.