Kling 3.0 Omni Image

Kling 3.0 Omni Image is the "latest model, all jobs" node from Kling: it generates images from a prompt, and it can edit existing images - the same node does both, which the description ("create or edit images with the latest model") doesn't exactly shout about. It also has a party trick the rest of the family lacks: generate a whole series of images in one call, so you can knock out variations without re-queueing. Partner node, ships with core, no install, cloud-hosted, billed per call.

The model pick is the headline. kling-v3-omni is the general do-everything model; kling-image-o1 is the reasoning-style model that thinks about the image before rendering - and comes with constraints: no series generation, no 4K. The node hard-errors if you ask for either, so it's not a silent gotcha.

How it works

Prompt goes to Kling's omni-image endpoint through Comfy's proxy; the node polls the task and returns the image(s) as a regular IMAGE tensor. Edit mode is prompt-driven rather than mask-driven - you attach one or more reference_images (up to 10), and your prompt describes what to change. The prompt field itself can carry both positive and negative descriptions, so you don't need a separate negative box.

The inputs that matter

• prompt - describes the content (or the edit), positive and negative in one box.
• model_name - kling-v3-omni or kling-image-o1. O1 is the careful, expensive one.
• resolution - 1K, 2K, or 4K. 4K doubles the price tier and is off-limits for kling-image-o1.
• aspect_ratio - 16:9 through 1:1 and back, including ultrawide 21:9.
• series_amount - disabled by default; set it to 2–9 to get that many images in one run. That's the batch feature, and it multiplies the bill by the series count (the badge reflects it). Not supported on kling-image-o1.
• reference_images - up to 10, used for editing or style/subject guidance.
• seed - same partner-node caveat as always: it controls re-running, not the outcome.

Output is IMAGE, batched if a series ran.

Where people get burned

• Mixing the two models' rules. kling-image-o1 + series = error. O1 + 4K = error. Both are raised before anything is billed, which is decent of it, but read the error before assuming the node is broken.
• Reference constraints. Uploaded references must be at least 300×300 and within a 2.5:1 aspect ratio band, and the node validates dimensions before calling. A tiny or ultra-panoramic reference gets rejected.
• Series cost. Asking for nine images "just to see" is a real line item. Series is for when you know you want variety, not for exploration.
• Prompt = edit instruction. If you attach references but write a plain generation prompt, you'll get a generation influenced by the references rather than a precise edit. Say what changed.