Krea 2 Large · Text to Image

Krea 2 was the model that took the open-weight ecosystem away from Ideogram in June 2026, mostly on licence terms and trainability rather than raw sharpness. What Krea sells above that open release is the hosted tier - and Krea 2 Large is the top of it: their most detailed model, priced at about 9 credits per image (nine cents) and aimed squarely at cinematic keyframes you'll then animate.

Why use the hosted tier at all

If you have the VRAM, you can run Krea 2 locally - the open weights correspond to the hosted "medium" tier, so the local file is a real substitute for the cheaper node. Large has no local equivalent, and that's the whole reason this node exists in a pack that otherwise asks "why rent what you can own". Nine cents buys one good keyframe without touching a sampler, a LoRA stack or a VAE choice, which for a lot of people is a fair trade.

Because Krea 2 is only open conditionally - commercial use free under $1M trailing revenue, license above - the hosted route is also the boring, safe way to use it commercially without reading a licence.

How it works

The node posts your prompt and settings to PopNow, which renders on Krea's infrastructure and returns images. The pack downloads them and stacks them into one IMAGE batch for your graph. There's no local diffusion step, no seed to hunt, and no sampler settings to tune - the model's own defaults do the work.

Inputs

prompt - subject, composition, style, lighting. That order, in that many words' worth of detail; Krea 2 responds well to plain descriptive writing rather than tag soup.

aspect_ratio - the widest choice of any node in this pack: 1:1, 4:3, 3:2, 16:9, 2.35:1, 4:5, 2:3, 9:16. The 2.35:1 cinemascope option is the reason to be on this node rather than Medium, which doesn't offer it.

reference_image_urls - an autogrow socket group, up to ten images that steer the style of the result. Wire one image and the next socket appears.

reference_strength - 0 to 1, default 1, and only meaningful when references are connected. At 1 the composition can carry over too; lower values keep mostly palette and texture. If your reference is someone else's artwork and you want the look not the layout, drop this to 0.3–0.5.

Outputs

images (IMAGE) and task id (STRING). The image batch goes into Save Image, a local upscaler, a masking pass, or a Kling image-to-video node to become a shot. The task id is for the API console.

Install

ComfyUI Manager → search PopNow → Install → restart. Or:

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

Then a key from popnow.ai/api/console/keys into Settings > PopNow, or POPNOW_API_KEY, plus one more restart so the node list reflects your account. Nothing gets downloaded: no checkpoints, no VAE, no CLIP. Delete custom_nodes/popnow_api.py if you ever installed the single-file node.

Notes from using it

Nine cents an image changes your habits. You stop generating a 30-image grid and start writing one careful prompt, which is arguably the right habit anyway. It also means batches are where costs hide - a batch of eight is 72 credits, $0.072, still trivial, but if you wire a Krea node into a loop over 200 prompts you've bought a small subscription.

If a render comes back with the composition right and the surface wrong, that's a prompt problem, not a model problem: name the material and the light source. If it comes back with the wrong framing, that's aspect_ratio plus reference images fighting your text - drop reference_strength before you rewrite the prompt.

Codes you may hit: 40201 (no credits), 401 (bad key), 400 naming a field (usually a combination of aspect ratio and something else), content_rejected when the model declines - refunded, so a retry with softer wording costs time only.