Recraft V3 (fal)

Recraft V3 is the design world's quiet favourite - the model that showed up on the Artificial Analysis leaderboard as the anonymous "red_panda," scored #1 with a 72% win rate over Midjourney v6.1 and Flux 1.1 Pro, and then held the top spot for five straight months. Its calling card is text-in-image accuracy and a huge curated style system aimed at commercial designers rather than hobbyists. This node lets you drive it from a ComfyUI graph instead of Recraft's own web app.

Like everything in the fal-API pack, it's a cloud call: no weights on your disk, no GPU used locally, and you pay fal per image. Recraft's weights are closed, so a hosted call is genuinely the only way to run it - there's no local checkpoint to download even if you wanted one.

How it works

You write a prompt, pick a size and a style, and the node ships it to Recraft V3 on fal. The model returns a finished image. The reason people reach for Recraft specifically over a local Flux or SDXL run is two things a generic checkpoint struggles with: legible typography (logos, posters, UI mockups, anything with real words in it) and a deep library of consistent, named art styles.

The inputs and outputs that matter

• prompt - multiline. Recraft is strong at following layout and text instructions, so be specific about any words you want rendered.
• style - this is the node's real lever, and it's a big enum. The families are realistic_image, digital_illustration, and vector_illustration, each with sub-styles like realistic_image/hdr, realistic_image/studio_portrait, realistic_image/motion_blur, or digital_illustration/pixel_art. There are 24 in total. Pick the sub-style that matches your job - it does more for the look than most prompt wording.
• image_size - presets (square_hd, portrait_16_9, landscape_4_3, and friends) or custom. If you choose custom, the width and height fields (512–2048, in steps of 16) take over.

There's also an optional style_id string, for feeding a saved Recraft style reference if you've built one.

Output is a native IMAGE, so it drops straight into a Preview Image, a Save Image, or any downstream node - no URL round-trip like the video nodes in this pack need.

How to install it

Ships in gokayfem's ComfyUI-fal-API pack. Via ComfyUI Manager: search ComfyUI-fal-API, install, restart. Or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/gokayfem/ComfyUI-fal-API.git
pip install -r ComfyUI-fal-API/requirements.txt

Then add your fal key: copy config.ini.example to config.ini in the pack folder and paste the key from fal.ai/dashboard/keys, or export FAL_KEY=your_key. Restart ComfyUI. Nothing else downloads - the model runs on fal.

Common issues & troubleshooting

The style isn't landing. The style enum outweighs your prompt for overall look. If you asked for "vector logo" in the prompt but left style on realistic_image, you'll get a photo. Set the sub-style to match.

Text comes out garbled anyway. Recraft is better than most at text, not perfect. Keep the words short, spell them exactly, and avoid cramming a paragraph into one render.

Custom size ignored. width/height only apply when image_size is set to custom. On a preset, the preset wins.

Insufficient funds / billing. fal bills per call. It's a legit paid platform; if a credit top-up hasn't shown up yet, let it settle before retrying rather than double-paying. Errors from fal (quota, content policy, validation) surface as readable text right in ComfyUI - read the message, it usually names the exact problem.