FLUX.2 Klein 4B · Edit

Klein 4B does generation and editing in one checkpoint, which is the whole reason it displaced a lot of Qwen-Image-Edit workflows in 2026 - you stop maintaining two models. This is the editing half, and at about 1.89 credits ($0.0189) an image it's the cheapest FLUX.2 edit in the pack. Wire a photo in, describe the change, get a photo back.

Not an inpaint node

Get this straight before you wire anything: the Flux 2 edit nodes here take reference images and instructions. There is no mask socket. If you've used the pack's FLUX.1 Fill nodes you're used to feeding a black-and-white mask; this isn't that. You describe what changes and what must stay, and the model decides where to touch. That's more forgiving than a mask when the change is semantic ("swap the jacket for a leather one") and much less obedient when it isn't ("change the number on the door from 14 to 22").

image_urls: the socket that grows

image_urls is an Autogrow input. You'll see one image socket; connect something and another appears, up to three. That ordering is meaningful - the model refers to sources by position, and the service numbers them image 1, image 2, image 3 in the prompt, so write your instruction with those positions in mind: Keep the woman from image 1, put her in the room from image 2, match the lighting of image 2.

One trap worth knowing about, because it fails at the API rather than on the canvas: each socket can carry a whole batch, and every image in the batch counts as a reference. Wire an IMAGE batch of four into one socket and you've just sent four references to a model that takes three, and you get a PopNow 400 naming the field. Send batches through a node that gives you one image, or split them first.

aspect_ratio defaults to auto, which keeps the shape of the first source image at roughly one megapixel - usually what you want for an edit. Choose an explicit ratio only when you're deliberately reframing. resolution is 1k or 2k; since the edit endpoints charge a flat rate rather than a size tier, the bigger one isn't costing you extra at the moment (the node's badge is the live answer).

prompt is the edit, in prose, up to 5,000 characters. Say both halves: what changes and what must not. And leave out the old habits - Flux 2's Qwen3 encoder ignores prompt weighting entirely.

seed starts fixed so an accidental re-queue replays the cached result instead of charging again. Move control after generate to randomize when you want a genuinely different attempt.

Outputs

images is a normal IMAGE batch - Save Image or a preview. task id is the PopNow job identifier; keep it for the API console and for support.

Install

ComfyUI Manager → Custom Nodes Manager → PopNow → Install → restart. Or:

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

Key from popnow.ai/api/console/keys into Settings > PopNow (or POPNOW_API_KEY in the environment for a server install), restart again so the node list is read from your account. The pack installs no pip packages - dependency list is empty - but it needs a ComfyUI with the V3 node API, so update before you file a bug about a missing node.

What it's good at, and where it stumbles

Small 4B models are confident, not careful. Ask for a jacket swap and it'll look great; ask it to preserve a specific face and you'll see drift, because Klein holds identity less tightly than Qwen-Image-Edit does. Two things help: name what must stay ("maintain the face and posture"), and keep the edit small. Chained micro-edits beat one heroic instruction.

Also expect the occasional warm shift in the result - a known Klein habit, not a colour-management bug in your graph. If it matters, fix it downstream by matching your result against the source.

If a run comes back as content_rejected, that's the hosted model declining, and no re-roll fixes it. Failed runs are refunded; refused ones are a different problem.