RecraftAI Image To Image Transformer

Image-to-image on a closed, server-side model - that's the whole pitch here. You feed this node an existing image and a text prompt, and Recraft's API transforms it into a new image that follows the prompt while staying recognizably related to the source. If you've used local img2img with a denoise slider, you already understand the shape of this; the difference is that the actual model lives on Recraft's servers, not on your GPU.

So why reach for it instead of a local checkpoint? Three reasons that overlap with the rest of this pack. First, it's Recraft's own model - the leaderboard-topping one behind "red_panda" - which means design-oriented restyling with genuinely strong text rendering. Second, zero local cost: no VRAM, no weights, nothing to download. Third, the ComfyUI-native workflow: the input and output are both plain IMAGE tensors, so this slots into a graph where the whole pipeline is local except this one step. That's the honest trade to weigh - you pay per call, but you only pay when you actually need the closed model's quality.

The inputs that matter

Required: client, image, prompt, and strength. Optional: style, substyle, seed.

The strength float is the dial that matters, and the node's default of 0.5 is a reasonable starting point. It maps to how far the output departs from the input, on the same 0–1 scale as a local denoise slider: low strength keeps the composition and just nudges the style, high strength lets the prompt rewrite the image. It's the first thing to touch when results feel either too conservative or too unhinged.

style and substyle work exactly like they do on the generator: realistic_image, digital_illustration, or any, plus seventeen substyles like studio_portrait, hdr, motion_blur, pixel_art, hand_drawn, and enterprise. For restyling this is where you actually steer - "make this photo a poster" is a substyle problem, not a prompt problem.

seed is a plain integer sent as random_seed. It behaves like any seed: fixed value, reproducible look.

How it works

The node serializes your IMAGE tensor back to PNG bytes, then multipart-POSTs it - along with the prompt, strength, style, and substyle - to Recraft's /images/imageToImage endpoint. The API returns a URL, the node downloads it, and converts it back into an IMAGE tensor on the single image output. Under the hood this is the same client object you get from the RecraftAI Client node, so the same key, billing, and gotchas all apply.

One thing to keep in mind: this is whole-frame transformation, not targeted editing. The prompt steers the entire image. If you want to change one region and leave the rest pixel-identical, that's the Inpainter node's job - see that article.

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 model downloads. You'll need an API key from recraft.ai/profile/api, and a positive API units balance to generate one.

Where people get burned

• Empty prompt = hard fail. The node raises Prompt is required. It won't silently do a passthrough.
• Expecting local-style control. You don't get to pick steps, sampler, or CFG. Strength is your only real lever, so changes feel coarser than local img2img. That's inherent to a closed API, not a bug.
• Cost. Every transform is a paid API call. Iterate on the prompt with a fixed seed rather than re-rolling the seed ten times.