Recraft V3 Text to Image

This is the hosted recraftv3 model as a node - the same one that spent five months at the top of the HuggingFace text-to-image leaderboard under the anonymous codename "red_panda" before the community figured out it was Recraft.ai. If that means nothing to you, here's the short version: it's a closed, design-leaning image model that beat Midjourney and Flux in blind rankings for a stretch, and its real strengths are illustration, clean composition, and consistent brand-style output.

You're not running it locally - you're renting it through ComfyUI's partner-node system. Give it a prompt, it spends your Comfy account credits on Recraft's GPU, and hands back a tensor.

The inputs that matter

• prompt - multiline, up to 1000 characters
• size - a combo from Recraft's size list: 1024x1024 default, plus portrait and landscape variants like 1024x1536, 1365x1024, up to 2048x1024
• n - 1 to 6 images per run
• seed - sets re-run behavior; results are nondeterministic regardless (their tooltip, and true)
• optional recraft_style - the RECRAFT_V3_STYLE from the style nodes. This is where Recraft's look actually lives; the default is plain realistic_image
• optional negative_prompt - "no text, no watermark" and so on
• optional recraft_controls - from the Recraft Controls + Recraft Color nodes, for color-palette steering

Output is a standard IMAGE tensor - wire it to Preview or Save like anything else.

How it works

Your settings are packaged into a request and POSTed to /proxy/recraft/image_generation with model recraftv3. The generated image comes back as a download, is decoded into a tensor, and you're done. If you ask for n images they're concatenated into one batch tensor.

Getting it

Ships with ComfyUI core - no extension, nothing to install. You do need to be signed into a Comfy account with credits on it (Comfy Cloud includes a monthly partner-node allowance).

Common issues

• The style system is the actual product. A bare Recraft generation with no style wired in is a fairly vanilla photoreal render - fine, but it's not why you'd reach for Recraft. Hook up a style node (Digital Illustration is the showpiece) and the model becomes dramatically more interesting. Most people's "Recraft looks mid" take is really "Recraft with default settings looks mid."
• Seed hunting is futile. There's no lucky seed to find; the API is nondeterministic regardless. You tune the prompt and style, not the seed.
• Cost per image. With n up to 6, one run can be six billings. It's cheap per image, but it adds up fast when you're iterating.
• This is the V3 node. If you came here looking for Recraft's newest - V4/V4.1 with its bigger sizes and better quality - that's the separate RecraftV4TextToImageNode. This one is the V3 workhorse and takes the style-object inputs V4 doesn't.

Reach for it when you want Recraft's design aesthetic without leaving the node graph, or when you need a hosted fallback your own GPU can't match. Just bring a style node with you.