AI Video Generation With Kling 3.0 — Hits and Misses

Why I Went Down This Rabbit Hole#

Saturday night, 11 PM. I had a product demo script that needed a 15-second hero clip — some sweeping aerial shot of a fictional city at sunset. No budget for stock footage, no time to learn Blender.

I opened my agent and typed: “generate a slow aerial flyover of a neon-lit cyberpunk city at dusk, cinematic.”

Ten seconds later, the skill had interpreted my request, routed it to the Standard text-to-video endpoint, and queued a job on RunComfy.

Three minutes later, I had a clip. It wasn’t cinematic. It was weird. But it was something.

Setup: How Long It Actually Took#

The skill itself installs in about 30 seconds — npx skills add against the agentspace-so monorepo with the --skill kling-3-0 flag. No API key to configure if you already have RunComfy set up. The harder part was figuring out which of the six endpoints to use for what.

The Kling 3.0 skill exposes Standard, Pro, and 4K tiers, each with text-to-video and image-to-video modes. Standard renders in about 90 seconds for a 5-second 1080p clip. Pro takes around 4 minutes for the same duration. 4K took 12 minutes for my first attempt and came back with 15 seconds of footage instead of 5.

The first three clips I generated were borderline unusable — characters with three arms, surfaces that rippled like water, lighting that shifted between frames. That’s not a setup failure, that’s the baseline prompt engineering cost. I spent about 45 minutes tweaking prompts before I got something I’d show someone without apologizing.

What I Built With It#

I needed a short brand video: a product shot transitioning through three stages — raw materials assembling, the finished device rotating, a lifestyle scene of someone using it.

I figured I could do this as three separate Kling 3.0 clips and stitch them.

Attempt one: I prompted Standard text-to-video for “precision mechanical assembly, top-down view, bright studio lighting.” The result was a gray blob that occasionally sprouted geometric shapes. Total miss.

I pivoted to image-to-video. I generated a still frame using gpt-image-2 of the product on a clean white surface, then fed that as the base image to Kling 3.0’s image-to-video endpoint. The result was dramatically better — the camera slowly orbited around the product, the lighting held steady, and the motion was smooth enough to use.

The failure-to-pivot arc taught me something: Kling 3.0’s image-to-video mode is significantly more reliable than text-to-video for anything involving specific objects. Text-to-video works for abstract scenes (weather, landscapes, particle effects). If you need a specific thing to look like itself, start with a reference image.

I tried the same workflow with Pro tier and noticed that motion quality improved noticeably — camera pans were smoother, object boundaries stayed cleaner — but the generation time tripled. For a 5-second clip, Pro took about 4 minutes versus Standard’s 90 seconds. The 4K tier was overkill for this project, but I tested it anyway: 12 minutes for a 15-second clip where the product stayed recognizable throughout.

What Surprised Me#

The 4K tier is not what I expected. I assumed 4K meant “sharper version of the same clip.” Instead, the 4K endpoint generates longer form content — 15 seconds instead of 5 — with noticeably more coherent scene structure.

The Standard tier gives you short loops. The 4K tier gives you actual scenes with a beginning, middle, and end. One of my 4K clips had a camera pan, a subject entering frame, and a lighting change across the duration. It was the first time I felt like I was watching a real video instead of an animated GIF.

The consistency between clips generated from the same prompt was also better than I expected. Running the same text prompt twice through Standard gave me two completely different interpretations — different camera angles, different compositions. Running the same image through image-to-video twice gave me near-identical motion paths. If you need reproducibility, start with an image.

Things That Annoyed Me#

The queue times are unpredictable. Standard tier clips come back in 90 seconds most of the time, but I had one instance where it took 11 minutes with no status feedback.

There is no progress indicator — the skill submits the job and you wait. You cannot cancel a queued job from the skill. If you realize you used the wrong prompt, you either wait for the result or restart your agent session.

The 4K tier failed silently on two out of eight attempts. The clip just never arrived. No error message, no timeout, nothing.

I only noticed because I was checking the RunComfy dashboard manually. The skill layer doesn’t surface these failures — it returns a success response when the job is submitted, not when it finishes.

And the character consistency problem is real. If your clip needs a person, expect their appearance to shift between shots. The same prompt generated a man in a blue jacket in one clip and a woman in a red dress in the next. Kling 3.0 has no concept of character identity across separate generations.

When I Use It vs When I Don’t#

I reach for Kling 3.0 when I need a quick visual concept — a mood board clip, a background video for a landing page, abstract motion for a presentation. The 4K tier is genuinely useful for short standalone scenes that don’t need character consistency. Pro tier at 1080p is the sweet spot for quality vs. speed.

I avoid it when I need specific, recognizable objects behaving predictably — unless I have a reference image to start from. And I never rely on it for anything with a hard deadline, because the queue variance and silent failures are too risky.

The skill does one thing well: it turns six endpoint configurations into one natural-language interface. If you accept the prompt engineering overhead and the queue uncertainty, it saves real time compared to working with Kling’s API directly. For quick visual concepts and mood boards, it’s become my default — just not for anything with a client deadline attached.