FLUX.1 Krea Dev LoRA · Edit

FLUX.1 Krea [dev] exists because Black Forest Labs let someone else tune their model. Krea - the San Francisco lab that ran a closed generation service for years before training anything of its own - applied its aesthetic to Flux in a July 2025 collaboration, and the result was the least "AI-looking" FLUX checkpoint of its generation. Skin with pores. Photographs that don't glow.

This node is that model doing an edit pass on your image, with a strength dial and a socket for up to three LoRAs, for 3.6 credits a run.

What it's good at, specifically

The complaint about base FLUX.1 Dev is well documented in the KB: homogenised faces, plastic skin, an over-lit look that makes everything resemble a stock photo shoot. Krea tuned against exactly that. If your goal is "looks like it was shot on a camera", this is the FLUX tier to start from - and there's a decent chance it's the right call over a newer, heavier model whose LoRA support is thin.

It's a Dev-family checkpoint, so guidance-scale logic applies: guidance-distilled, no negative prompt, guidance_scale default 3.5, range 0-20. That's the dial that governs the plastic look, and dropping it to 2.0 is the perennial trick for faces that look like they were generated rather than photographed.

The fields

image_url is the source, connected as an IMAGE. Result keeps its aspect ratio.

prompt - describe the whole image you want back, not just the change, and put any LoRA trigger word first. This node's tooltip adds the trigger-word note explicitly, which tells you it expects your own weights.

strength defaults to 0.8, range 0 to 1. That's img2img territory, not subtlety: 0.8 will re-render the photograph in a way that keeps its composition and subject rather than preserving pixels. For restoration-ish work come down to 0.4-0.5.

resolution is 1k or 1.5k on the long side. output_format is jpeg, png or webp - use png if you're going to grade afterwards, since jpeg on top of a photographic model is two kinds of loss.

seed starts fixed. Queueing the same graph twice reuses the first result rather than billing a second.

loras takes up to three files, each with its own strength, fed from a PopNow LoRA node: a .safetensors in ComfyUI/models/loras (uploaded once, cached 12 hours), a Hugging Face repo id or page, a Civitai page, or a direct https:// link. Gated files need HF_TOKEN or CIVITAI_API_KEY in ComfyUI's environment - the pack keeps tokens out of widgets because widget values travel inside shared workflows and PNG metadata.

Note what's absent: no step count, no num_outputs, no mask. This is a one-image-in, one-image-out restyle. If you want masking, the Dev LoRA edit node and the Fill node in the same pack have it.

Outputs are images and task id.

Under the hood

The node's inputs come from black-forest-labs/flux-1-krea-dev-lora/edit's published schema. On queue it uploads your image to a popnow:// URI, quotes the job before holding credits, submits under an idempotency key, polls, downloads and decodes into an IMAGE batch. The status line finishes with the actual charge.

Install

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

Or ComfyUI Manager, search PopNow, install, restart. Key from popnow.ai/api/console/keys into Settings > PopNow or POPNOW_API_KEY, then restart again so the catalog is rebuilt. No GPU, no model files, no extra packages.

Where people get burned

Krea's realism is a look, not a truth serum. It sharpens a good photograph and it will also faithfully reproduce a bad one's problems - noise, bad colour, a blown highlight - because you asked for 0.8 strength and it redrew at 0.8 strength.

Second: the LoRA pile labelled "Krea" on Civitai is a mix of FLUX.1 Krea [dev] fine-tunes and Krea 2 ones, which are different architectures entirely. Krea 2 is a later, bigger, separately trained model with its own settings; a Krea 2 LoRA won't do anything useful here. Read the base model field on the download page before you spend a run on it.