Kling 3.0 Image

Kling made its name on video, but the same Kuaishou lab ships stills too, and this node is the stills machine: Kling 3.0 Image turns a text prompt into an image through Kling's hosted API - optionally with a reference image that it treats far more carefully than most cloud generators do. It's a partner node, so it lives in ComfyUI core with zero install, runs entirely in the cloud, and bills per call against your Comfy account.

The genuinely interesting part is the reference-image handling. Most text-to-image nodes let you throw a reference at them and pray. This one splits the question in two: image_fidelity ("how much should the uploaded reference bleed into the result") and human_fidelity ("how closely should a subject's appearance match"). If you've ever fought an img2img workflow for consistency, you'll appreciate that these are separate dials.

How it works

Text prompt goes to Kling's image-generation endpoint via Comfy's proxy; the node polls until the task finishes and downloads the result as a regular IMAGE tensor. If you connect a reference image, it's uploaded alongside and the two fidelity sliders tune how strongly it's used. Pick kling-v3 or the older kling-v2 as the backbone, choose an aspect ratio, and queue.

The inputs that matter

• prompt / negative_prompt - the positive and negative sides, both multiline. Kling's image model responds to plain descriptive sentences.
• image (optional) - your reference. Without it, this is a plain text-to-image node; with it, you get subject/character guidance.
• image_fidelity - 0 to 1, default 0.5. Reference intensity for the uploaded image. Crank it up when you want the reference to dominate.
• human_fidelity - 0 to 1, default 0.45. Subject-reference similarity - this is the one for keeping a face or person consistent.
• model_name - kling-v3 or kling-v2. V3 is the current one and costs about twice as much per image; the price badge shows it.
• aspect_ratio - default 16:9, with the usual portrait/square options.
• n - up to 9 images per run. Nine images, nine charges - the badge does the math for you.
• seed - the standard partner-node caveat: controls re-running, not reproducibility. Results are nondeterministic regardless.

Output is IMAGE, batched when n > 1, wired straight into Save Image or further processing.

Where people get burned

• Seed is a lie, sort of. The tooltip says it plainly: results are nondeterministic no matter the seed. Don't treat it as a reproducibility lock.
• Reference quality in, reference quality out. The fidelity sliders can't rescue a blurry or tiny reference. The node's guidance is generous but it still needs a decent source to hold a subject.
• Cost stacks with n. It's tempting to ask for nine variants in one run; each one is billed separately. Generate one, look, then iterate - you'll spend less.
• Negative prompt isn't always enough. It's there and it works, but don't expect ControlNet-grade exclusion. If you need precise composition, you're in the wrong node family.