FLUX.2 Klein 4B · Text to Image

Klein is German for small, and Klein 4B is the smallest thing Black Forest Labs has shipped in the Flux 2 line - four billion parameters, Apache 2.0 weights, four-step distilled, and roughly a second per image on a good card. Through this node it's about 1.4 credits an image, which makes it the least expensive way to find out whether an idea is worth rendering properly.

That's the use case, and it's a good one: this is a draft machine, not a delivery tool.

Why the small model earns its place

There's a whole local argument about Klein 4B - it fits in about 8.4GB of VRAM on a 5090 run, it's the first commercially licensed Flux 2 checkpoint, and it does editing as well as generation. None of that matters here, because the compute isn't yours. What matters is the price ratio: at ~$0.0139 a shot you can explore twenty compositions for less than a quarter, then take the winner to Klein 9B or Dev.

The known weakness travels with the model. Klein's anatomy is its reputation: extra fingers, occasionally a third arm, unusual poses falling apart. On the local side the standard fix is running 8 steps at CFG 1.2–1.5 instead of 4 at CFG 1.0. You cannot do that here - there's no steps field and no guidance field on this node, because the endpoint owns them. So the hosted equivalent of that trick is: generate a few and pick, or prompt for simpler poses, or crop and inpaint the hands afterwards. Which, honestly, is what you were going to do anyway.

Fields

prompt - up to 5,000 characters, and worth using. Flux 2's encoder is a Qwen3 LLM, so full sentences with clear ownership beat comma tags: A woman in a linen shirt sits on a windowsill, late afternoon sun across her face, shallow depth of field. Weighting syntax is dead (((like this:1.4)) does nothing), and so is the old masterpiece, best quality ritual.

aspect_ratio - 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 21:9, 9:21. Pick the delivery ratio now rather than cropping later; you're paying by the image, not by the pixel.

resolution - 1k (about a megapixel) or 2k (2,048px long side). The 2k step costs a little more; for drafting, 1k is plenty.

seed - starts fixed. That's a billing guard: the same graph queued twice gives you the same picture instead of a second charge. Switch control after generate to randomize when you want fresh takes.

Outputs

images is a proper IMAGE batch - Save Image, Preview Image, or straight into a local upscale/refine chain if you're using this as a first pass. task id is a string for the API console and support.

Install

ComfyUI Manager → Custom Nodes Manager → search PopNow → Install → restart. The git route, if you prefer to see what you're installing:

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

Then generate a key at popnow.ai/api/console/keys, drop it into Settings > PopNow, and restart one more time - the node list is built from your account's catalog at startup, so skipping that second restart is why people think a model is missing. Zero pip dependencies; the pack uses only what ComfyUI ships, but it needs a ComfyUI new enough for the V3 node API.

Small print and small traps

Don't wire a Load Image batch anywhere near this node's text-to-image sibling by accident - text-to-image takes no image input at all, and a stray socket means you've grabbed the wrong node. Also worth knowing: the node's search aliases include the raw model id, so typing flux in the canvas search box finds a dozen PopNow nodes at once. Their display names are the reliable guide - "· Text to Image" versus "· Edit".

Last thing: 1.4 cents a picture is cheap, but a graph that loops over 40 prompts isn't. Put a number in Settings > PopNow > Max credits per run and you'll never learn this the expensive way.