Custom Recraft API Node

Recraft is the model line you reach for when legible text in an image actually matters - typography, lettering, vector-style illustration - the stuff diffusion models usually mangle into mush. Custom Recraft API Node is the pack's preconfigured doorway to it: an image-generation node whose request payload is shaped for Recraft, pointed by default at an OpenAI-compatible aggregator that serves it.

The honest framing first: this node doesn't call Recraft's official API. It POSTs to the api_url you give it using the same images/generations shape the pack's other image nodes use. The default URL is HyprLab, the author's aggregator of choice. If you have a key for a provider that serves recraft-v3 - or the model name your endpoint actually uses - swap the URL and key and it just works. That's the whole philosophy of this pack: the node is glue, the provider is your choice.

How it works

It's a single requests.post with a JSON body built from your inputs - model, prompt, size, and style (that's where the style_type dropdown maps to - the payload field is literally called style), plus response_format: "url" and output_format: "webp". It expects the standard OpenAI-style response envelope, reads data[0]["url"], and returns that URL. tenacity retries on connection errors and timeouts, three attempts, so a hiccup won't nuke the run.

The inputs that matter

• prompt - what you want drawn.
• model - default recraft-v3; change it if your provider names it differently.
• size - a 15-option dropdown of resolutions, from 1024x1024 up through 2048x1024.
• style_type - the 23-option style dropdown: digital_illustration and its children (pixel_art, hand_drawn, grain...), realistic_image variants (b_and_w, hard_flash, studio_portrait...), and vector_illustration (engraving, line_art, linocut). These are Recraft's own style vocabulary, and they're the whole point - pick vector_illustration and you get the clean flat look the model is known for.
• seed - reproducibility.
• api_url / api_key - the provider and its key. Defaults are HyprLab + YOUR_API_KEY_HERE, so both need attention.

Output: image_url (STRING). Note that - it's a URL string, not an IMAGE tensor. To get pixels on a wire you chain it into a load-image-from-URL node (the pack ships one, "Load Image from URL (BillBum)", and it also accepts data URIs if your provider returns those instead).

Installing it

Part of AhBumm/ComfyUI_BillBum_Nodes - ComfyUI Manager search "billbum", or:

cd ComfyUI/custom_nodes
git clone https://github.com/AhBumm/ComfyUI_BillBum_APIset_Nodes.git
cd ComfyUI_BillBum_APIset_Nodes && pip install -r requirements.txt

Restart ComfyUI. Deps are light - requests, pillow, numpy - no models to download. This is an API node; the "model" lives on the server, not on your disk.

Where people get burned

• Default URL vs. your key. HyprLab is the default; if you set it to a generic OpenAI-compatible endpoint, remember that plain GPT-image-style endpoints won't understand Recraft's style field. This node is tuned for providers that speak Recraft-ish payloads.
• Style dropdown is provider-dependent. A random images endpoint may ignore style silently. If output looks generic, your provider isn't honoring Recraft styles.
• Output is a URL. Don't wire image_url straight into an img2img node expecting IMAGE - convert first.
• Recraft's calling card is text rendering; if that's not what you need, the pack's other image nodes are less opinionated about style.