Recraft V3 Image to Image

Image-to-image is the same trick you know from local ComfyUI - start from an existing image instead of noise, steer it with a prompt - but the diffusion happens on Recraft's servers, not your GPU. You get the recraftv3 model's particular look, which for this node's audience is the whole point: Recraft is the design-focused model that quietly sat at the top of the text-to-image leaderboard for months under the codename "red_panda," and its image editing is popular precisely because it handles the illustration/design side of things so well.

The one input that matters more than the rest is strength, and its tooltip is gloriously honest: it "defines the difference with the original image, should lie in [0, 1], where 0 means almost identical, and 1 means miserable similarity." That's exactly how it behaves. Leave it at 0.5 and you'll get a confident reinterpretation. Push toward 1 and the source becomes a vague suggestion; drag it toward 0 and the prompt barely dares to whisper.

The inputs that matter

The full schema is small enough to memorize:

• image - what you're transforming
• prompt - multiline, what you want
• strength - 0.0 to 1.0, default 0.5, step 0.01
• n - 1 to 6, how many results per input image
• seed - mostly cosmetic, see below
• optional recraft_style - feed it a RECRAFT_V3_STYLE from the Recraft Style nodes to lock in a style family and substyle
• optional negative_prompt - what you don't want
• optional recraft_controls - from the Recraft Controls node, for color/background control

Output is a standard IMAGE tensor, so it plugs into Preview, Save, or whatever you'd do with any generated image. Batch in, batch out - each image in your input batch is run through the pipeline.

How it works

Your image is sent up to /proxy/recraft/images/imageToImage alongside the prompt and settings, and the result is pulled back down. That's the whole mechanism - hosted inference through ComfyUI's own proxy, billed per image from your Comfy account credits. No local VRAM, no model download. This is one of those cases where "ships with ComfyUI core" means "ships with core but needs a Comfy account and credit balance."

Common issues

• Seeds are theater. The tooltip on the seed input says it outright: it only determines whether the node re-runs; actual results are nondeterministic regardless of seed. Don't hunt for a "lucky seed" like you would with a local sampler - you'll chase your tail.
• Strength reads differently across styles. Recraft's style system is strong, and a heavy substyle will overpower the source image faster than a plain realistic look. If edits are drifting too far, drop strength before you touch the prompt.
• Style consistency. Unlike local img2img where you control the checkpoint, here the model is fixed - the way to get consistent looks across a run is to reuse the same style node and settings, not to tweak seeds.

If your only goal is a quick recolour or light retouch of something you already like, you could do it locally for free. Reach for this node when you want Recraft's design/illustration aesthetic applied on top of an existing image - that's its home turf, and it's very good at it.