Kling O3 Review: Reference Casts and Video Editing, Filtered at L2

We don't publish our content-ladder runs; what follows is the everyday register, taken from the published, safe-for-work examples on the Kling O3 model page.

Two women under a shared umbrella. An opening frame of two women sharing a transparent umbrella on a neon-lit street, with a prompt asking only for heavier rain, a laugh and a walk away from camera — their faces matched the source frame through the whole clip.

A sealed letter on a night train. Three reference images — a man, a 1930s dining carriage, a wax-sealed letter — combined into one shot: the man sitting alone at a table, breaking the seal, a scene none of the three photos showed on its own.

A multi-shot tango. A text-only prompt asking for a wide shot, a close-up and a low angle came back as three cuts in that order, holding wardrobe and street lighting steady across all three.

Winter over Lisbon. A video-edit job turned a sunny bike-ride clip into deep winter — fresh snow, wool coats, cold blue light — while the riders, the yellow tram and the framing carried over from the source clip unchanged.

On image-to-video and video-edit, the source frame or clip fixes who's in the shot and what they're wearing; write the prompt around what changes, not what stays. On text-to-video with shot_type: multi, list the shots you want in order and Kling will cut between them while holding the rest of the scene steady.

1. Multi-shot from text alone.

Wide shot, then a close-up, then a low angle. A night tango on a rain-slick cobblestone street in Buenos Aires. A man in a dark waistcoat and a woman in a backless black satin dress dance close under a single streetlamp, rain beading on the stone.

Run this prompt on Kling O3 →

2. A change, not a rewrite, for video-edit.

Turn the season to deep winter: fresh snow on the cobblestones, the balconies, the rooftops and the yellow tram, snowflakes drifting through the air and cold blue morning light. Dress both riders in wool coats and knitted scarves. Keep everything else the same.

Run this prompt on Kling O3 →

Weak: describing a whole new scene on video-edit invites the model to redo more than you asked for. Strong: name exactly what changes and add "keep everything else the same," the way prompt 2 does — it's what kept the tram and the riders' faces intact.

Kling reviews every request itself, at three points: the prompt text, any uploaded image, and the finished clip. That's true across all four O3 tasks — the same outcomes came back on every route we tested, so this is Kling's own moderation rather than something a different endpoint would change.

On 28 September 2026 we ran our five-rung ladder against all four O3 endpoints. L1 Suggestive (lingerie, a slipping strap) rendered on every task. L2 Partial was refused every time, on every task, including six further L2 probes across six O3 and Kling 3.0 endpoints the same day. L3 Strong was refused on all three O3 tasks we sent it to — text-to-video, reference-to-video and video-edit — in under a minute when it came as an uploaded photo or clip, but only after about two minutes on text-to-video. We didn't run L4 Intense (an intimate scene between two adults) or the L5 Extreme restraint-and-impact probe on O3. The L5 Extreme gore probe split by task: it rendered on image-to-video and was refused on reference-to-video.

Refused requests cost nothing — every refusal in our test set, including the ones that only came back refused after the clip had fully rendered, was free. There's no quiet softening either: in every run the clip either came back as written or the job failed outright, with clothing never added back and the scene never swapped for a safer one. Retrying a refused video-edit job won't help; a rejected source clip returns a generic error, not a reason.

The page label reflects this: all four Kling O3 tasks carry a filtered policy tier, and we add no content filter of our own on top of it.

Not tested yet: L4 Intense and the L5 Extreme restraint-and-impact probe on any O3 task; L3 Strong on image-to-video; scenes with more than two people at the harder rungs; reference-to-video with more than three photos at the upper rungs.

Adults only, in every prompt, upload and output. Never depict a real, identifiable person without their consent, and make sure you hold the rights to any photo or clip you upload. For content past lingerie level, use one of our Spicy families instead — Wan 2.2 Spicy and Seedance 2.5 Spicy are two of the video ones — this endpoint refuses that far past L1.

Kling O3 doesn't have much of an independent following outside the wider Kling conversation, but it gets named directly when developers talk about closed-source video tools. In an April 2026 Hacker News thread, one commenter wrote: "The models most professionals use are Kling o3, Kling 3.0 and, more recently, Seedance 2.0".

Public blind-vote boards back that up only loosely: on Artificial Analysis's image-to-video leaderboard (28 September 2026), Kling 3.0 Omni's 720p Standard tier placed 21st of 82 tracked models, a little behind Kling 3.0's own placement on the same board.

• Need motion transfer from a performance clip instead of a reference photo? Kling 3.0 adds Motion Control, driving a character photo from a 3-to-30-second clip, at the same core prices as O3's image and text tasks. Our Kling 3.0 review.
• Don't need reference casting or video editing? Kling 2.6 covers image and text-to-video with native audio from $0.069 a second, without O3's extra tasks.
• Want dialogue-heavy scenes to be the main focus? Seedance 2.5 is priced higher but built around lip-synced speech.
• Need a single continuous take longer than 15 seconds? Wan 3.0 runs up to 30 seconds with multi-shot direction, from $0.045 a second at 480p.