Kling 3.0 Omni Transformation

What it does

Video in, video out. You hand it a clip and a prompt - "turn this into a cinematic night scene while keeping the characters' movements" - and it re-renders the clip as a new clip that keeps the motion and changes the look. Style transfer for footage, plus the ability to fold still images into the same job.

MiniMax H3's launch chatter was full of people describing V2V motion transfer "like Kling" as the reference standard for this. Transformation is the KIE-hosted Kling 3.0 Omni route to that class of result, straight from a ComfyUI graph.

Standard framing for this family: Kling is Kuaishou's closed model with no downloadable weights, so the API call is the only path, and here it's KIE.ai - an account and credit balance of its own, not ComfyUI's official Partner Nodes. KIE_Next_kling_3_0_omni_transformation_77531962, KIE Next / Video / Kling.

The inputs

• prompt - required, up to 3072 characters, and it's the whole job description. Say what should change; the motion comes from your footage.
• video_1, video_2, video_3 (VIDEO, optional) - your source footage. Three numbered sockets exist because the underlying API field is an array. You'll normally use one.
• image_urls (IMAGE, optional) - still images folded into the same generation, which is how you get "same scene, but this product/character instead".
• duration - a plain text field, default 5. The tooltip says this is configurable when you're using a video with images, so leave it alone on a pure video-only run.
• resolution - 720p, 1080p, 4k.
• aspect_ratio - 16:9, 9:16, 1:1, auto. And the constraint is genuinely useful: auto is only valid for video-only input. The moment you add images, you must pick a real ratio. This is exactly the kind of thing to get wrong at 2am.
• audio - off by default.
• elements - the one-time subject assets array, same as the text-to-video sibling (up to seven image subjects, up to three video-character subjects, with reduced caps when mixed).
• timeout_seconds / callback_url - 1200 seconds of polling; leave callback empty.

What happens under the hood

Your VIDEO input gets materialized to an MP4 and uploaded to KIE's file host; images are PNG-encoded and uploaded the same way. The node then POSTs kling-3.0-omni/transformation to createTask and polls until the task reports success, growing the poll interval from ~2.5s to a 15s cap. The finished file is downloaded, archived under output/KIE-Results/videos/, and returned as a native Comfy VIDEO - which means you can feed it straight into a local upscaler, a frame extractor, or a Save Video node without any re-encoding dance.

The one thing to internalise: this is a synchronous job with a network round trip and a render queue behind it. Ten seconds of footage is not instant, and the timeout widget is there so you don't click Generate twice.

Outputs

video for the graph, url for the KIE-hosted link, all_urls_json, task_id, raw_json (your failure post-mortem), plus credits_consumed and credits_left. Both credit figures also print on the node after each run. credits_left shows -1 if KIE's balance endpoint is unavailable; the generation isn't affected.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/felipederosilva/ComfyUI-KIE-Nodes-Next

Restart, then Settings → KIE.ai Nodes Next → Connection → Paste / replace KIE.ai API key. Validation happens before the key is stored, the frontend only ever sees a masked fingerprint, and you get a ✅ CONNECTED status with your balance. Server installs can use KIE_API_KEY, which takes priority. Requirements are only requests, Pillow, numpy, PyYAML - this pack does not drag in torch extensions or video codec libraries of its own. Manager install works when the pack is listed; the repo also publishes a ZIP and a Windows one-click installer under dist/.

Where it goes wrong

• auto aspect ratio with an image attached. Rejected by the API. Pick 16:9 or friends.
• Long source clips. The reference-video field on the motion-control sibling documents a 3–30 second window; treat anything much longer as a gamble and cut first.
• Stacking this on a fresh install. The pack generates its model nodes from KIE's live docs index, so a brand-new install shows the bootstrap set first and the rest after a sync and restart. Missing node ≠ broken install.
• Assuming cheap. Every run is metered, video-to-video is a heavy job, and iteration is where money goes. Queue one, read the credit receipt, then decide how you feel about a sweep.
• Expecting a local bypass. Outputs are moderated at the provider, and nothing about installing this pack changes that.