RecraftAI Creative Upscaler

This is the sibling to the Crisp Upscaler, and it's the one with actual imagination. Where crisp just makes things bigger and cleaner, Recraft's "creative upscale" boosts resolution with a focus on refining small details - and, per the README, faces specifically. That's a generative upscale in the "more detail" bucket of the community's standard taxonomy: the model isn't just interpolating pixels, it's inventing plausible detail that was never there.

Which means the honest warning up front: this will change your image. The upscaling essay in the knowledge base puts it bluntly - generative upscaling "will rewrite faces." Recraft's version is tuned to make faces better (sharper eyes, cleaner skin texture) rather than drifting into uncanny, but it's still generation. The eyes, the hairline, the fabric weave - all of it gets inferred, and on a 100%-zoom comparison the output is not the input.

So when do you reach for it? Two cases. One: the source is soft or low-res and you want a plausible, detailed version - Recraft's model trained for this. Two: you're already in a RecraftAI workflow and want the face-refinement pass without installing a local pipeline like SUPIR or SeedVR2. The KB's verdict on those local options is that they're the current gold standard for detail-add; this node's edge is that it's one click, runs server-side, and is explicitly tuned for portrait detail. Its weakness is the same as every closed API: no scale control, no steps, no local fallback, and you pay per call.

The inputs and output

Required: client (from the RecraftAI Client node) and image (any IMAGE tensor). Optional: seed.

There's no scale factor, no strength slider - the API decides how much to upscale. Feed it a portrait and the face gets the extra attention. The output is an image tensor at higher resolution with refined detail, ready to save or composite into the rest of your graph.

How it works

The node serializes your image to PNG, multipart-POSTs it to Recraft's /images/creativeUpscale endpoint with your API key, downloads the result, and converts it back into a ComfyUI IMAGE tensor. Same client, same key, same billing model as the rest of the pack - server-side compute, nothing runs locally.

Installing it

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

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

Windows portable: .\python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\ComfyUI-RecraftAI\requirements.txt. Only requests, torch, Pillow, numpy - no models to download, no VRAM used. You'll need an API key from recraft.ai/profile/api with a positive API units balance.

Where people get burned

• Assuming it's a clean upscale. It's not - it's generative. Always compare the output at 100% zoom against the source before trusting it in a chain. If the subject's identity matters (a real person, a client's product), a generative pass can drift features you can't afford to lose.
• Wrong tool for clean sources. If the image is already sharp and detailed, creative upscaling adds nothing but changes everything. Use the Crisp Upscaler for that job instead.
• No size control. You get Recraft's default output scale, not the 2x/4x you might want. Plan your graph around whatever comes back.
• Cost. Paid per call, and since it's generative, re-rolling the seed to pick the best face costs you per attempt. Set the seed once and judge honestly.