Kling Image Generation

Kling Image Generation is the pack's text-to-image workhorse: type a prompt, get a Kling v3 image back as a normal IMAGE tensor, exactly where a local sampler would sit in your graph. It's the image-side complement to the video nodes, and it's the node you reach for when you want Kling's specific look - the same model family that powers its video - without leaving ComfyUI.

The honest framing first, because it shapes everything: Kling's image model is closed, hosted, and metered. There's no weights file to download and no VRAM involved - your prompt goes to Kling's servers and comes back as a result, billed in credits. The community's standard objection to API generation ("it's not local, data leaves the machine, it costs per call") applies fully here. What you get in exchange is a capable hosted model with no local install, which is exactly the use case the KB's API-wrapper pattern describes: a model you were never allowed to download, reached through a node that looks like any other generator.

How it works

The inputs that matter:

• auth - the KLING_AUTH object.
• prompt - multiline; your main instruction. This is the one you'll spend most of your time on.
• negative_prompt - things to avoid, plain and simple.
• model_name - currently just kling-v3 (the dropdown has one entry as of this version).
• aspect_ratio - one of the seven Kling ratios, default 1:1.
• resolution - 1k (~1024px) or 2k (~2048px). 2k costs more credits; pick by destination.
• fidelity - a 0..1 slider, default 0.5. The tooltip says it plainly: 0.0 = creative, 1.0 = faithful. Lower values let the model be looser and more interpretive; higher values push prompt adherence. It's the Kling equivalent of a CFG knob, and 0.5 is a sane middle.
• n - how many images in one call, 1 to 9. Batch up when you're iterating; Kling returns them as a stacked IMAGE batch.

Outputs: image (IMAGE - a batch if n > 1), url (the hosted image URL), and task_id. The image output wires straight into anything that consumes images: save nodes, VAE-ish pipelines, or as the start of an Image-to-Video flow.

How to get a good one

• Write for a hosted filter. Kling enforces its own content policy server-side - there's nothing to abliterate, no weights to patch. If a prompt gets refused, it's refused; rephrase or route around it.
• Use fidelity as your creative dial. Stuck in bland output? Drop fidelity toward 0.3 and give the prompt room. Need exact adherence (product shots, text)? Push toward 0.7+.
• 2k is a real cost jump. Use 1k for drafts, 2k only for the final that ships. The Cost Estimator node can show you the spread before you commit.
• Chain it into video. The natural pipeline is Image Gen → Image-to-Video, giving you a consistent Kling look through the whole generation.

Common issues

• Rejected prompt - content policy or a malformed parameter. Check the error string; if it's moderation, rephrase.
• n > 1 returned fewer than requested - Kling sometimes returns fewer images than asked for; the node raises if none come back, but a partial batch is possible. Treat n as "up to n."
• Aspect ratio surprises - pick the ratio deliberately. Kling won't letterbox; a mismatch with your downstream video ratio means cropping later.

Install

Ships in ComfyUI-Kling-Direct:

cd ComfyUI/custom_nodes
git clone https://github.com/IxMxAMAR/ComfyUI-Kling-Direct

or ComfyUI Manager → search "Kling Direct" → install → restart. No models, no GPU, no extra deps - just the libraries ComfyUI already bundles. If you're already paying for Kling video, this node makes the images in your pipeline match it.