Krea 2 Turbo LoRA · Text to Image

What this node actually is

Krea 2 Turbo, the distilled 12B model Krea opened in June 2026, running on PopNow's servers, wrapped up as one node. The interesting part isn't the base model - it's the loras socket hanging off the bottom. You point a PopNow LoRA node at a .safetensors file, connect it here, and your style file gets applied to a cloud render. No 12B checkpoint on your disk, no 8-step sampling loop, no VRAM math.

Worth knowing why anyone cares. Krea 2 shipped a genuinely undistilled base (Raw) alongside Turbo, and the community filled the LoRA library within days - Comfy said on the joint Krea livestream that he'd never seen a modern base released without an aesthetic finetune. Turbo is the 8-step, CFG 1.0 version of that model, so this node is roughly "Krea 2 Turbo at someone else's electricity," plus your own adapter on top.

Reach for it when you've trained or adopted a Krea 2 style you like and don't want to babysit it locally, or when your card can't hold a 12B at a comfortable resolution. It is not the cheapest way to make a Krea image - that's a local 8GB FP8 run - but at around two cents per 1k image it's not the hill to die on either.

How it works

The node is a thin, honest client. Your prompt and dropdown settings go to POST /api/v1/jobs/quote first, then a task is created and polled until the file comes back. Anything you connect to a media socket gets uploaded for you and swapped for a popnow://f/... reference only your key can use, and identical uploads are cached for 12 hours, so tweaking the prompt doesn't re-send the same picture. The result is decoded into an ordinary ComfyUI IMAGE tensor.

Krea 2 itself is a single-stream DiT with Qwen3-VL as the text encoder and the Qwen-Image VAE - which is exactly why a Krea 2 LoRA is the only kind that will do anything here. Adapters are architecture-bound; an SDXL or Flux file in that socket is dead weight at best.

The inputs that matter

• prompt - the author's own note is the right advice: describe the picture in sentences, and put the LoRA trigger phrase first. Long natural language, not tag soup.
• resolution - 1k is about a megapixel, 2k about four, both generated natively rather than upscaled. It's the field that moves the price.
• aspect_ratio - fifteen options from 1:1 through 21:9 to 1:3. Pick it here; there's no crop afterwards.
• seed - starts on fixed deliberately. Queue the same graph twice and you get the cached first result rather than a second bill. Switch control-after-generate to randomize when you actually want variation.
• loras (optional) - wire in the PopNow LoRA output. Most LoRA models take up to three, so chain nodes to stack.

Outputs are images (a single IMAGE, plural name notwithstanding - straight into Preview Image or Save Image) and task id, a string you paste into the API console or hand to support when something goes sideways.

Install

The pack installs clean - its pyproject.toml declares no dependencies at all, because it uses only what ComfyUI already ships.

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

Or in Manager, search PopNow and hit Install; the CLI equivalent is comfy node install popnow-comfyui. Restart, then grab a key from popnow.ai/api/console/keys and paste it into Settings > PopNow (or set POPNOW_API_KEY before launching, which is the right move on a server). Restart once more after the key is in - the node list is built from your account's catalog at startup.

Where people get burned

The LoRA node with an empty source adds nothing, silently. It's not an error; there's just no adapter in the stack, and you'll spend ten minutes blaming the trigger words. Fill in the source or leave the LoRA node out of the graph entirely.

Gated or private files need HF_TOKEN (Hugging Face) or CIVITAI_API_KEY (Civitai) in ComfyUI's environment. Tokens are read from the environment only - never from a widget, because widget values get baked into shared workflows and PNG metadata. And files over 2 GiB are refused outright.

Two errors worth recognising: PopNow 40201 means the wallet, not the graph (buy credits or drop to 1k); PopNow 40004 means the model can't serve that exact combination and the message names the field to change. Refused prompts come back as content_rejected and are refunded, which is the one consolation of Krea's famously twitchy filters. Set Max credits per run in Settings if you'd rather a misclick stop before it holds anything.