Krea 2 Turbo LoRA · Edit

Cloud models are usually a closed box: you get the prompt, you get the output, you don't get your own style. This node breaks that - it takes a source image, a prompt, and a loras input you feed from a PopNow LoRA node, so a .safetensors file from your ComfyUI/models/loras folder can steer a hosted Krea 2 render. From about 3.24 credits per image, roughly a third of a cent per attempt after that.

How it works

Two nodes, one job. The PopNow LoRA node resolves a source - a local LoRA file, a Hugging Face repo or page, a Civitai page, or a direct https:// link - and hands a POPNOW_LORAS value to this node's loras input. PopNow uploads or imports the weights once (up to 2 GiB, cached for 12 hours per key) and applies them to the Krea 2 Turbo render. You can chain several PopNow LoRA nodes; most models accept up to three.

The image side is a classic edit: your image_url goes up as PNG (JPEG fallback above 10 MiB), and the model returns a re-rendered image that keeps the source's aspect ratio.

Inputs

prompt - and this is the one people get wrong: describe the whole image you want back, not just the change. This is an img2img-shaped model, not an instruction editor. "A woman in a linen shirt standing in a sunlit kitchen, warm morning light, 35mm" gets you a good edit; "make her shirt linen" gets you mush.

image_url - the source image. Aspect ratio carries through; resolution is chosen separately, so don't expect the source's pixel dimensions.

resolution - 1k (about one megapixel) or 2k (about four). This is the main price dial.

strength - 0 to 1, default 0.6. How far to move from the source. Low values preserve structure and detail; high values let the LoRA and prompt rebuild the image. For a style LoRA over a photo, 0.45–0.65 is the useful band.

seed - an integer. Same seed and settings, same result; every run is billed, so control after generate starts on fixed and queueing twice reuses the first result. Switch it to randomize when you want variations, and remember each one costs.

loras - the POPNOW_LORAS socket. LoRAs must be trained on Krea 2 Raw or Turbo; a Flux or SDXL LoRA will not do anything useful here.

Outputs

images (IMAGE) and task id (STRING). Straight into Save Image, a batch upscale, or onward to a video node.

Install

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

Restart, paste a key from popnow.ai/api/console/keys into Settings > PopNow or set POPNOW_API_KEY, restart again so your catalog becomes the node list. Manager users can search PopNow. Nothing else installs - no models, no dependencies. Delete any old custom_nodes/popnow_api.py.

For gated downloads, set HF_TOKEN or CIVITAI_API_KEY in ComfyUI's environment. Tokens are read from the environment only, never from a widget - deliberately, because widget values get saved into every workflow and every PNG you share.

Where it breaks

Local LoRA files: source takes the file name as it appears in ComfyUI/models/loras, and the upload happens once on first use. A 2 GiB cap is generous for a LoRA and short for a full checkpoint, so don't paste a merged model in and expect it to work.

Krea 2's own quirks travel with the hosted model. The open weights shipped with a trained-in alignment pass that flattens expression and drops prompt clauses about bodies and violence; hosted tiers are a different build, but you should still expect refusals on edge-case content, and a content_rejected run is refunded rather than charged.

And the boring ones: 40201 empty wallet, 401 dead key, 400 naming a field when a setting is rejected, 40310 if your email isn't verified. Cancelled runs keep rendering on PopNow, get charged if they succeed, and land in the API console - check there before you queue the same thing again.