arkennemasis fal · Flux Krea Image to Image · $0.025/MP

Same Krea model as the text-to-image node, different entrance: you hand it a picture and a prompt, and it re-renders that picture with the Krea look. $0.025 per megapixel, no local weights. This is the node you want when the composition is already right and only the rendering is wrong - mushy skin, flat light, that unmistakably AI sheen.

How img2img actually behaves here

The node uploads your image to fal's CDN, then fal re-noises it according to strength and denoises it back under your prompt. Low strength keeps almost everything you gave it and just re-points the detail; high strength lets the model invent. Now the interesting part: the default here is 0.95, and the tooltip says outright that higher strengths are better for this model. On Krea, "0.95" doesn't mean "barely touch it" - it means the model is doing nearly all the work while borrowing your layout and palette. If you were hoping for a subtle 0.3 pass, this isn't the endpoint for it. Crank it down to 0.5–0.7 and you'll mostly get the original back with different grain, which is fine, just not what people expect from the number.

Note what's missing: there is no image_size widget on this node. Output follows the input image's shape. If you need a specific aspect ratio, crop or pad before the node rather than hoping.

The inputs worth setting

• image - a normal ComfyUI IMAGE. A batch goes through as-is; feed it one picture per node if you want per-image settings.
• strength - 0.01–1, default 0.95. The single most important control.
• prompt - describe the result, not the input. "Same scene, shot on a Fuji, late afternoon" beats "photo of a street".
• num_inference_steps - default 40 here (10–50), higher than the text-to-image node's 28. Leave it.
• guidance_scale - default 4.5, range 1–20.
• seed - -1 for random. The node returns whatever seed was used on its seed output.
• enable_safety_checker, output_format (jpeg/png), acceleration (none/regular/high).

Then the pack-wide four: max_cost_usd (default 20, 0 = no cap), reuse_identical_run (on - it will serve you the exact same paid-for result again instead of re-billing, unless you turn it off for a fresh take), extra_json for fields this node has no box for, and max_concurrent.

Outputs

images goes to a preview or save node. image_paths shows where they landed under output/fal/fal-ai_flux_krea_image-to-image/. timings, info and has_nsfw_concepts are there for debugging and for knowing whether the safety checker ate your result. The prompt output echoes what was sent, which is handy when an upstream text node builds your prompt for you.

Install

ComfyUI-Manager → search comfyui-arkennemasis, or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/Hishamahmer/comfyui-arkennemasis

Restart afterwards. Add your fal key to the .env file in the folder that holds run_nvidia_gpu.bat (portable) or in your ComfyUI folder:

FAL_KEY=your-key-from-fal.ai/dashboard/keys

The node appears at arkennemasis/fal/Image/Flux. There's no key field on it by design, so your key stays out of the workflow JSON you share.

Where it goes wrong

Black images. The usual cause is the safety checker flagging the output. It comes back as a black frame rather than an error, and has_nsfw_concepts tells you that's what happened. Disabling the checker requires fal account authorization; unauthenticated requests are checked anyway.

Wrong aspect ratio. As above, there's no size widget. Pad or crop upstream.

Surprise on the second run. If you re-queue the same node with the same input image and settings, reuse_identical_run gives you the files from the first run for free. That's the point - but if you were trying to sample a different take, set it to false or change the seed, otherwise you'll be staring at the identical output wondering why.

Four images, four times the price. num_images can go to 4 and the badge multiplies accordingly. Watch the number on the node before you queue.