Qwen Image (fal.ai)

This node is the shortest route to Alibaba's Qwen-Image without owning the hardware. Qwen-Image is the 20B diffusion transformer that made text rendering look easy - crisp glyphs, clean posters, believable UI mockups, and it handles Chinese as well as English, which almost nothing else does. The catch is that 20B of MMDiT wants a serious GPU, so most people never ran it locally at all. FAL_QwenImage just phones out to fal.ai's hosted fal-ai/qwen-image endpoint, waits, and hands you the result as a normal ComfyUI image. Your GPU basically takes the call off while a server farms it out; you pay per image instead.

That's the whole deal with this pack - it's a wrapper, not a model. MIT-licensed, tiny, and cleanly typed by elle-trudgett (look for ComfyUI fal.ai Nodes in ComfyUI Manager). The heavy lifting happens inside fal.ai, which is worth remembering when the image finally lands: it's cloud compute on their tab, and the meters are running.

How it works

The node builds a request dict from your settings and calls fal.subscribe("fal-ai/qwen-image", arguments) through the pack's API handler. fal runs the generation, and the node downloads the result images and converts them back into a batched IMAGE tensor. From the outside it behaves like any local sampler node - wire it into a Preview or Save Image and you're done.

The two options that actually change what you get:

• prompt - the only required input. Qwen responds well to natural-language descriptions; it's a prompt-adherence heavyweight, not a tag-parser.
• image_size - a preset (square_hd, square, portrait_4_3, portrait_16_9, landscape_4_3, landscape_16_9) or custom, in which case custom_width/custom_height take over. Default is landscape_4_3, which is a slightly odd default for a first run - switch to square_hd or your target aspect before you burn a few calls on the wrong shape.

After that: num_inference_steps (default 30), guidance_scale (default 2.5 - Qwen likes CFG down around here, don't crank it like SD1.5), seed (-1 = random), an optional negative_prompt, and output_format (png/jpeg). The acceleration and use_turbo toggles trade quality for latency on fal's side; leave them off for the first pass.

The output is one IMAGE - a batch if you set num_images above 1 (which multiplies the bill, by the way). Feed it into anything that eats images: an upscale pass, a Qwen-Image-Edit node from this same pack, a Save Image.

Installing it

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 use ComfyUI Manager and search the pack title. No model files download - there's nothing local to download. Then give it a key:

export FAL_KEY="your-api-key"   # from fal.ai/dashboard/keys

ComfyUI Desktop? Drop a config.ini into the fal-ai-comfyui-nodes folder with [API] / FAL_KEY = your-key. Restart ComfyUI either way.

Common issues

• "FAL_KEY not set" - the classic first-run wall. The node reads the env var, then config.ini; set one, restart, done.
• Black or obviously-flagged outputs - the enable_safety_checker toggle defaults to true, and fal's filter can return a flagged image when it decides your prompt crossed a line. You can turn it off, but their content policy still applies to the account.
• Silent cost - this is a paid API. A quick settings tinkering session can rack up a surprising tab, and it's easy to forget the node isn't local.
• Errors surface as fal.ai error: ... - the pack unwraps fal's message for you; a rate-limit or balance error usually tells you exactly what's wrong.

For fast local work on a 24GB card, Z-Image is lighter and free; reach for this node when you want Qwen's specific strengths - text, posters, that 20B adherence - without the VRAM.