抠抠图-通用放大变清晰功能

KoukoutuUpscale (抠抠图-通用放大变清晰功能) is the "make it bigger and cleaner" node of the comfyui-koukoutu pack, and it does the job without loading a single upscale model on your machine. You feed it an image, pick a factor, and it phones the Koukoutu cloud API, which upscales and restores the image server-side and hands back the result. For anyone whose workflow is pinned to a 6GB card, that's the whole pitch: SUPIR and SeedVR2 are great, and they're also VRAM hogs that rewrite your faces if you look at them wrong. This is the hosted version of the "more detail" job from the upscaling playbook - the tradeoff is you pay a little credit per call and your image leaves your machine.

The honest positioning matters here. If all you want is more pixels on an already-clean image, a local ESRGAN model or Lanczos is free, instant, and can't hallucinate - reaching for a hosted upscaler for that is paying for nothing. This earns its keep when you want genuine restoration (de-blur, de-noise, detail rebuild) in one node with no model management, or when the source is destined for print and you need a clean 4x or 6x without tiling. The pack's own example workflow even chains it right after a cutout (upscale2stamp) - cut the product out, then upscale it clean for the print file.

How it works

Async, like most of the pack: upload the image, create a task (model key upscale2stamp), poll once a second with a ComfyUI progress bar, download the result URL when done. Server 5xx errors retry up to five times, and there's a five-minute cap before it gives up.

Inputs that matter

• image - any IMAGE tensor. First frame only if you feed a batch.
• api_key - required; get it at https://www.koukoutu.com/user/dev and keep credits topped up.
• scale - 2, 4, or 6, default 4. 4x is the sweet spot for most product work; 6x takes longer and is mostly for print-size output.
• skip_error - on by default.

Outputs

image plus a message string ("成功" on success). Same fallback as the other async nodes: with skip_error on, an error returns your original, un-upscaled image plus the error text on message - so "the node did nothing" usually means "read the message output."

Install

It comes with the whole pack - install once, get all nine nodes. In ComfyUI Manager, search "koukoutu". Or:

cd ComfyUI/custom_nodes
git clone https://github.com/zhenzi0322/comfyui-koukoutu
pip install -r requirements.txt   # requests, Pillow, numpy, torch

Restart ComfyUI; it lands under image/koukoutu. No model files to download, dependencies are trivial, and your VRAM is untouched - the tradeoff is it's a paid API call and your image goes off-box. Install it from the official repo; it's an API wrapper that carries a key by design.

Troubleshooting

The 40x family is your first stop when it silently fails: 401 bad key, 403/409 invalid key or no credits, 406/413 file over 15MB, 407 image under 70px (too small to upscale meaningfully), 415 bad format, 422 bad params, 429 rate limit. The default 4x is usually right; if a small source upscales to mush, that's a source problem, not a settings problem - restore quality tops out at what the source actually contains.