Which models take a start frame, and which take both

Keyframing is the most precise control the video engine offers, and whether you get one anchor or two depends entirely on which model you picked.

What does a keyframe actually do?

Anchors the generation to an image you supply rather than to a description.

A start frame fixes where the shot begins. The model generates forward from a real image instead of inventing an opening from your prompt.

An end frame fixes where it finishes. With both supplied, the model synthesises the motion, lighting shifts and camera path between them — so you've defined the shot's two endpoints and asked it to fill the middle.

That's a different kind of control from a reference. A reference says "this is what the character looks like." A keyframe says "this is what frame one looks like." The second is much more specific.

A reference describes what's in the shot. A keyframe defines a specific frame of it. Only twelve models let you define two.

Which models take which?

All eighteen, from the picker.

Three things worth pulling out.

Vidu Q3-Turbo's whole identity is keyframing. Its description is "first and last frame, up to sixteen seconds," and it's the only model on the roster whose positioning is about the technique rather than about speed, quality or references.

The five start-only models cluster. Four of them are also the lowest-reference models on the roster — Hailuo 2.3 and Kling 3.0 Turbo take one reference each, Wan 2.6 and Grok Imagine 1.5 take three. So the models with the least reference capacity also tend to have the least keyframe capacity.

Gemini Omni Flash has no frames section. It also has no duration control and no resolution control. Its whole panel is references, prompt, aspect ratio, sound — which makes it the model with the fewest decisions available anywhere on the roster.

Where this falls short. The table says what a model accepts, not how well it interpolates. Two anchors don't guarantee a good path between them, and nothing in the picker tells you which models handle a long interpolation better than a short one.

Why does the difference matter?

Because a two-anchor workflow fails quietly on a start-only model.

If your process is "generate the opening frame, generate the closing frame, let the model connect them," switching to HappyHorse 1.1 or Kling 3.0 Turbo mid-project doesn't produce an error. It produces a shot generated forward from your start frame, ending wherever the model decides.

Which may look fine in isolation and won't match what comes after it.

The models most likely to catch you out are the fast tiers, because those are what people switch to when iterating — and Kling 3.0 Turbo is start-only while Kling 3.0 Omni takes both. Same family, different capability.

Switching to a start-only model doesn't error. It just quietly ignores half your plan.

What are keyframes actually good for?

Three things, and the third is underused.

Anchoring a shot to a specific image. A generated still you're happy with, a photograph, a frame from an earlier clip. Anything you can produce as an image can become frame one.

Defining a transformation. Start frame is the room lit, end frame is the room dark. The model generates the change rather than you describing it, which is more reliable than prompting "the light fades."

Carrying continuity across a cut. This is the one people miss. Take the last frame of a shot, use it as the start frame of the next one, and the second shot begins exactly where the first ended — same light, same positions, same space. That's the most precise continuity tool the engine has, and it doesn't require a prompt to describe anything.

How does this interact with Extend?

They solve the same problem from opposite ends, and it's worth knowing which to reach for.

Extend continues a rendered clip. The model has actual footage to continue from, so space, light and identity carry automatically. Seven models support it.

A start frame anchors a new generation to a single image. You get a fresh shot that begins where you specified.

So Extend is for making one shot longer, and a start frame is for making the next shot begin correctly. They're not alternatives — a sequence assembled through cuts wants start frames, and a single long take wants Extend.

Where this falls short. A start frame taken from the tail of a previous clip is one image. Extend continues from the clip itself, which carries motion the image doesn't. For anything where movement has to flow across the join, Extend is the better tool.

How should you use them?

Four steps, and the first prevents the failure at step 9.

All frame support read from Hexcoded's Creative Studio picker in Video → Create mode in September 2026, all 18 models observed individually. Capabilities change with model versions and differ by mode. Verify in the picker before relying on this.

More on model capability, access and rights in Models.