Kie Image API X

kie.ai is a reseller that proxies Google's Gemini image models (and others) at volume pricing and multi-region endpoints - the community route to Nano Banana and friends when you can't or won't set up a full Google Cloud billing account. Kie Image API X is WorkflowX's wrapper around that service: pick a model from its catalog, type a prompt, and the generated image comes back on a normal IMAGE socket.

The model catalog is the headline - it spans closed and open model families across the reseller's catalog: wan-2-7 / wan-2-7-pro, grok-imagine, gpt-image-1-5, gpt2, seedream-4-5, seedream-5-lite/pro, nano-banana-2/pro, qwen2, and flux2-pro. If you want to A/B Nano Banana against Seedream against Grok in one workflow, this is a cheap way to do it without juggling three vendor SDKs.

The node is asynchronous by design: it submits the job, then polls the provider until the image is ready. timeout_seconds (default 600) and poll_interval_seconds (default 3) control that loop, and retrieval_mode switches between generate and force_retrieve - the latter for picking up a previously-submitted pending task. Image generation over an API can take a while; the polling is how the node survives that.

The inputs that matter

• api_key - optional override; otherwise it reads KIE_API_KEY from the environment. Credential-free by default, which is deliberate - you can share workflows without leaking keys.
• model - the catalog choice, which changes which controls make sense.
• prompt, aspect_ratio, image_size (1K/2K/4K) - the generation basics.
• quality (basic/low/medium/high) and num_inference_steps (default 28) - the quality dials.
• guidance_scale (default 3.5), seed_enabled + seed - reproducibility controls; seed is off by default.
• nsfw_checker, enable_safety_checker, watermark - safety and provenance toggles. Remember the reseller still passes your prompt to the underlying vendor's model, which enforces its own moderation.
• enable_web_search, enable_image_search, thinking_mode - provider extras, useful on the models that support them.
• custom_size_enabled + custom_width/custom_height (64–8192) - when the preset sizes aren't right.
• show_payload - dump the actual request payload for debugging.
• Optional image_1 (IMAGE) - a reference image; the frontend adds more sockets as you connect them (up to fourteen).

Output: image (IMAGE) - the returned generation, ready for your normal downstream nodes.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/haroonaslam/WorkflowX-Configurator

or search WorkflowX Configurator in ComfyUI Manager, restart, refresh. No pip deps, no models - the "models" live on the reseller's servers.

Honest gotchas

• Reseller = middleman. You're paying for someone else's access to a model you can't run locally. Cost per call, inputs leave your machine, and "looser filtering" claims are vendor marketing until proven - the underlying model still refuses what it refuses.
• Timeouts are real. A 600-second default timeout exists because API generation genuinely takes that long sometimes. If you're getting timeouts, raise it or drop resolution/quality before blaming the node.
• Polling means the node runs long. In a batch, each API call blocks its own branch until the result arrives. That's normal for this pattern, not a hang.
• Model selection changes the visible controls - that's the frontend adapting to the catalog, which is helpful but means some toggles only matter on certain models.

The one-stop reseller node for people who want closed-model access without the billing-and-region friction, and the natural companion to the Atlas node on the same pack.