Qwen Image 3 Text to Image

Qwen-Image became a local favorite because it renders text and follows Chinese prompts better than almost anything open, and it runs on consumer GPUs. But the newest line, Image 3.0, is API-only - no weights, no local run, which is a notable shift for the team that made its name shipping Apache 2.0 checkpoints. This node is how you use it anyway: a partner node that calls Qwen-Image 3.0 through Comfy's proxy. Same family as the local Qwen, zero VRAM required, and your GPU is free to keep doing something else while the cloud renders.

What it is. A text-to-image call to the qwen-image-3.0 or qwen-image-3.0-pro hosted models, built into ComfyUI core. You don't install anything; you need a Comfy account with a credit balance, and each image is billed against it. No model files exist to download - that's the whole point.

How it works. The model dropdown is a dynamic combo: pick a model and it unfolds into the real controls - a prompt (English or Chinese), a negative_prompt (a rarity these days, and it actually does something), and width/height. Sizes run 256 to 2560 in steps of 16, with the total area clamped between 512x512 and 2560x2560, so any aspect ratio inside that area works. n generates up to 6 images at once, returned as a batch.

The inputs that matter. For a beginner, it's mostly the model combo - model choice, prompt, and width/height. The two toggles worth knowing: prompt_extend (default ON) has the model rewrite and enrich your prompt before generating, which helps short vague prompts but means what you typed isn't exactly what got used; and watermark adds an AI-generated marker if you want it. seed and n are what they always are, with the reminder that you pay per image - n=6 is six billings, not one.

The outputs. One: IMAGE, a batch of n results ready to wire into the usual preview/save/compositing nodes. No final_prompt here, so if you want to inspect what prompt_extend did, you can't - turn it OFF if you want deterministic control over what the model saw.

How you get it. Ships with ComfyUI core, no extension pack. Partner-node requirement in one line: log into your Comfy account, have credits, and your prompt (and nothing else) travels to the Qwen API. All these partner nodes landed in the mid-2026 wave, so keep ComfyUI current.

Where people get burned. The biggest confusion is treating this as the free local Qwen - it isn't, it's metered, and the pro model bills at a higher rate than the base one (the node shows a price badge so there's no surprise at checkout). Second: prompt_extend rewriting your carefully-worded prompt, which usually improves results but can break precise instructions - flip it off for exact control. And don't expect Image 3.0 quality barring from a 256x256 area; stay near the 1024 sweet spot unless you know what you're doing. If you genuinely want local, uncensored Qwen, the open Qwen-Image weights and their normal local nodes still exist and work great - this node is specifically the hosted path.