RecraftAI Image Generator

This is the headline act of the whole pack: a text-to-image node that runs Recraft's model - the one that held the number one spot on Hugging Face's text-to-image leaderboard for five months under the anonymous codename "red_panda." It beat Midjourney v6.1 and FLUX 1.1 Pro on that leaderboard and then stayed there while everyone tried to figure out who'd built it. That's the model this node calls, and you can reach it from inside ComfyUI without ever touching Recraft's web app.

But get the expectations right before you start. Recraft is design-focused, not photoreal-first. Its reputation is built on clean stylized output, very strong text-in-image rendering, and a huge style catalog - the official spiel leans on "long texts, brand styles, brand colors." If your job is "glossy editorial photo," local FLUX or an SDXL photoreal checkpoint is probably a better fit and costs nothing per render. If your job is "poster with a headline baked in," this is a genuinely different tool than anything you can run locally.

The inputs that matter

Required: prompt (multiline), and a client from the RecraftAI Client node. You can't run this without the client - see the pack install notes for the token setup.

The two inputs that actually shape the output are style and substyle. Style has just three real choices - realistic_image, digital_illustration, or any - and substyle is the finer dial with seventeen options like studio_portrait, hard_flash, hdr, motion_blur, pixel_art, hand_drawn, and enterprise. Leave both empty and you get the model's own judgment, which is decent but not the point of the node. The point is pinning a brand look: digital_illustration + hand_drawn_outline is a completely different animal from realistic_image + studio_portrait.

Also on the table:

• image_size - sixteen fixed presets from 1024x1024 up to 2048-wide and 2048-tall. No arbitrary sizes; pick the closest preset.
• model - recraftv3 or recraft20b. V3 is the flagship; 20b is the older/larger base.
• seed - an integer, sent to Recraft as random_seed. Set it to reproduce a look, or leave it and get a different result every run.

How it works

The node POSTs your prompt and options to Recraft's /images/generations endpoint with your API key as a bearer token, gets back a URL to the finished image, downloads it, and converts it into a normal ComfyUI IMAGE tensor on the image output. You can wire that straight into a save node or into further processing. The whole thing happens server-side - your GPU idles, your disk never fills up with a model, and every render bills you API units.

Installing it

ComfyUI Manager: search "RecraftAI" and install ComfyUI-RecraftAI, then restart. Manual install:

cd ComfyUI/custom_nodes
git clone https://github.com/recraft-ai/ComfyUI-RecraftAI
cd ComfyUI-RecraftAI && pip install -r requirements.txt

Windows portable uses .\python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\ComfyUI-RecraftAI\requirements.txt. Dependencies are just requests, torch, Pillow, numpy - no model downloads. You still need an API key from recraft.ai/profile/api, which requires a positive API units balance.

Where people get burned

• Empty prompt. The node hard-fails with Prompt is required - no silent blank images.
• Style blindness. style and substyle only do something meaningful together. digital_illustration with no substyle is vague; pairing it with one substyle is where the magic is.
• Cost surprises. There's no local fallback and no free tier caching. Generate ten seeds to pick one and you've paid for ten generations. Set your seed and iterate on the prompt instead of spamming seeds.