抠抠图-图生图功能

KoukoutuImageToImage (抠抠图-图生图功能) is the one node in the comfyui-koukoutu pack that does classic image-to-image: you give it a starting image and a prompt, and the Koukoutu cloud model generates a new version of it. No checkpoint to download, no VAE, no scheduler arguments - just image in, prompt, image out. If your whole reason for avoiding local img2img is that you don't want to babysit a diffusion stack, this is the zero-setup version.

Context helps here. The KB's inpainting essay is blunt that mask-free "edit the whole image" models have taken over most of what people used to do with img2img - and that's exactly the genre this node is: it takes the whole image into context and emits a new one. That's great when you want a restyle or a reinterpretation, and it's the wrong tool when you need the rest of the frame to stay bit-identical. For the latter, stay local with real masked inpainting. And worth saying plainly: img2img is the most consumable job in this pack - the other nodes do focused chores (cutout, print extraction, upscale), this one burns credits on creative exploration, so it's the easiest one to spend on.

How it works

Async, like most of the pack: upload your image, submit a task (model key image-to-image) with your prompt and settings, poll once a second with a progress bar, download the result. Server 5xx retries up to five times; five-minute cap.

Inputs that matter

• image - the starting image.
• api_key - required; koukoutu.com/user/dev with credits.
• prompt - required, multiline. Say what the new image should be.
• negative_prompt - optional, the usual "what not to do".
• similarity - 0–1, default 0.80. This is the dial that matters: higher means the result stays closer to the original, lower means the model runs further away. Think of it as the inverse of a denoise strength - 0.9 barely moves, 0.5 is a real reinterpretation.
• type - 0 or 1: the tooltip calls it "二次创作程度" - how far the remix goes. 0 is a light rework, 1 is a heavy one. Start with 0 and only step up if the change feels too timid.
• skip_error - on by default.

Outputs

image plus message. Standard pack fallback: with skip_error on, a failure returns your original image and the error text on message.

Install

Ships with all nine pack nodes; install once:

cd ComfyUI/custom_nodes
git clone https://github.com/zhenzi0322/comfyui-koukoutu
pip install -r requirements.txt

Or search "koukoutu" in ComfyUI Manager and restart. Light dependencies, no models, no VRAM. It's an API wrapper - your image and prompt go to the Koukoutu servers and are subject to their moderation, same as any hosted image model. Install from the official repo.

Troubleshooting

401 bad key, 403/409 no credits, 406/413 file over 15MB, 407 source under 70px, 422 bad params, 429 rate limit. If the output ignores your prompt entirely, similarity is probably cranked too high - drop it toward 0.7 and try again. And because this is a hosted model, there's no abliterating its content filter; what it refuses, it refuses.