Qwen3 Pro Image to Image

What it actually is

This is a cloud edit node wearing local clothes. You wire an IMAGE in, the node uploads it to KIE.ai, calls the qwen3/pro-image-to-image endpoint with your prompt, waits, and hands back an IMAGE tensor like nothing happened. No model weights, no VRAM, no sampler. The name is honest - it really is a Qwen3 image-to-image model behind an HTTP call - but the experience is nothing like the local Qwen nodes you may have installed.

Why you'd want that: instruction editing is what ate 2026. Feed a model one good picture plus a sentence and you get the same character in a new pose, new outfit, new scene, because the model is looking at the reference instead of re-inventing it. That's the default advice now, and local Qwen-Image-Edit is the usual tool. The catch is that it re-emits the whole frame, so pixels you never asked about move and identity drifts across a chain of edits. This node is the other trade: you lose local control, you gain a strong hosted model that runs on a laptop with integrated graphics.

How it works

KIE.ai is a reseller in front of a pile of closed and open models - the same category as the fal, Replicate and Runware wrappers, priced pay-as-you-go instead of subscription. The pack turns KIE's documented API catalog into individual nodes at startup, so this node exists because KIE publishes a page for qwen3/pro-image-to-image.

Under the hood: your image batch is uploaded to KIE's upload host and the URL is swapped into the request body; the node then submits a task and polls until the result lands. timeout_seconds (default 1200) is how long it's willing to wait before giving up - the job keeps costing credits either way, so don't set it to 30 and expect a fast model.

The inputs that matter

• image_urls - the reference. It's a batch socket, and KIE's docs cap it at three images. Don't feed it a batch of eight and assume extras are dropped quietly.
• prompt - the edit instruction. Both Chinese and English are supported.
• prompt_extend - on by default and I'd leave it. The author's own tooltip says the model rewrites your prompt and it "significantly improves results for simple descriptions." Turn it off when you're chasing an exact phrasing.
• resolution and image_size - 1K/2K crossed with eight ratios. Both are real money multipliers.
• output_format, nsfw_checker (off by default), negative_prompt, seed.

Outputs are the usual pack spread: image, url, all_urls_json, task_id, raw_json, credits_consumed, credits_left. Wire image into your Save/Preview; keep url if you want the copy that lives on KIE's CDN.

Install

ComfyUI Manager → search the pack title KIE.ai Nodes Next (manual clone is the reliable route today, since the README describes registry publishing as prepared rather than done):

cd ComfyUI/custom_nodes
git clone https://github.com/felipederosilva/ComfyUI-KIE-Nodes-Next
pip install -r ComfyUI-KIE-Nodes-Next/requirements.txt

That requirements file is four light libraries - requests, Pillow, numpy, PyYAML. No torch, no compile step. Restart ComfyUI, then paste your key once at Settings → KIE.ai Nodes Next → Connection. It validates the key before saving, shows a masked fingerprint, and keeps the key out of workflow JSON entirely.

Where people get burned

The upload limit bites: images over about 10 MB per file get rejected before generation starts. Long jobs hit the timeout - raise timeout_seconds rather than re-running, because a re-run bills twice. And KIE retires models: community threads have people discovering an endpoint they'd built around simply stopped existing. That's exactly what the pack's catalog sync is for - refresh it, and if this node's underlying model is gone, the request fails rather than silently changing behaviour.

Two KIE node packs exist and tutorials mix them up. If a guide's node names don't match what you see in the menu, you're reading about the other one.