Qwen Image 2 (fal.ai)

Here's the thing about Qwen-Image 2: the community has been begging for its weights since it launched in February 2026, and Alibaba kept them. It's a 7B unified generation-and-editing model with native 2K output - genuinely interesting architecture, far lighter than the 20B v1 - and there's no checkpoint to download. FAL_QwenImage2 is one of the few places you can actually run it, because it wraps fal's hosted fal-ai/qwen-image-2/text-to-image endpoint.

So this node is the "sneak peek" of the Qwen lineup, in a good way. You get the newer model's look and its 2K ceiling without the 20B VRAM tax, and the whole thing costs you a few API calls instead of a download. The trade: it's a rental, and it behaves like the new architecture - no steps, no CFG, no sampler. You prompt, it generates.

How it works

The node sends your prompt to fal's endpoint with the settings below, waits, and converts the returned images into a normal IMAGE tensor. That's the whole mechanism - fal.subscribe behind a tidy wrapper, one output, nothing local.

The inputs that matter:

• prompt (required) - the only required field. With no diffusion dials, the prompt is everything.
• image_size - presets (square_hd default, square, the portrait/landscape ratios) or custom, where custom_width/custom_height apply. Custom here is bounded to 512–2048px, which is the native 2K design - don't expect the 14K ceilings of the v1 node.
• enable_prompt_expansion (default true) - the model rewrites/expands your prompt internally before generating. This is the closest thing to a quality lever this node has; turn it off if you've already written a detailed prompt and want it taken literally.
• enable_safety_checker (default on), negative_prompt (optional, only sent when non-empty), num_images (1–4), seed (-1 random), output_format (png/jpeg/webp), private_mode.

Output: one IMAGE, straight into a Preview, Save, or the rest of the graph.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/elle-trudgett/fal-ai-comfyui-nodes
pip install -r fal-ai-comfyui-nodes/requirements.txt

or ComfyUI Manager → ComfyUI fal.ai Nodes. No model files to download - that's the point of an API wrapper. Set your fal key: export FAL_KEY="...", or a config.ini with [API] / FAL_KEY in the node folder for ComfyUI Desktop; keys live at fal.ai/dashboard/keys. Restart after.

Gotchas

• "FAL_KEY not set" - env var or config.ini, restart, done.
• No steps or guidance knobs. If you're used to tuning diffusion settings, this node will feel sparse - that's the model's design, not a bug. The prompt and enable_prompt_expansion are your levers.
• It's a paid API. A settings sweep on a metered 2K model adds up; iterate on the prompt first.
• 2K is the ceiling, not a guarantee - and higher sizes bill more.
• Errors arrive as fal.ai error: ..., unwrapped from the API so you can read them.

If you've been following the Qwen-Image 2.0 story and want to try the future before the weights (if ever) drop, this is the node. Just know you're renting it.