H3 Max Middle Frame: Plan a Three-State Transition

H3 Max can use a middle image, but only through fal's minimax/h3-max/reference-to-video endpoint. The image is a visual target near a chosen time. It is not a guaranteed cut, a frame-exact pose, or a feature of the ordinary H3 Max and H3 Max Turbo image-to-video routes.

That distinction is the whole workflow. A useful middle frame says, “Make this moment readable during the transition.” It does not say, “Nothing may resemble this image before frame 60.”

Three images should describe one shot

Start, middle, and end frames have different jobs.

The best three-image set looks like three stills taken from one plausible piece of motion. Keep the subject, camera height, lens feel, background geometry, and light direction recognizable. Change the state that matters.

For a product reveal, that might mean a closed package, the package half unfolded, and the finished display. For a character turn, it could be a front-facing pose, a readable profile, and a back-facing pose. If every picture contains a different scene, the model has to invent a montage when you asked for a transition.

This is different from supplying three ordinary reference images. References describe appearance or identity. The middle image has a time attached to it.

The field is a target, not an edit marker

The official schema calls the time field middle_frame_time. It is measured in seconds from the beginning of the requested clip, must fall strictly inside the duration, and is rounded to the nearest frame at 24 fps. A five-second clip therefore cannot use 0 or 5 as its middle time. 2.5 is valid.

The name sounds more exact than the generative behavior. In a published Voyager comparison, the same three flower images, prompt, and seed were generated with the middle target at two seconds and then four seconds. Voyager also supplied the middle flower as a subject reference in both requests. That extra conditioning may help explain why the flower state appeared early, although the only changed field between the two requests was the target time. The later version was already visibly open before the four-second target and held the intended state rather than waiting until that instant to enter it.

That is a useful result, but it is Voyager's experiment—not ours—and it should not be generalized into a frame guarantee. Use the field to move emphasis earlier or later. If a logo, legal line, or product label must appear on one exact frame, add it in an editor after generation.

There is another time unit worth keeping separate. H3 Max Camera Controls describes trajectory time on a normalized 0 to 1 scale. middle_frame_time uses seconds. Copying a value from one contract into the other changes its meaning.

Pick the middle image last

It is tempting to begin with three attractive pictures. A better process starts with the action.

Write one sentence that describes the whole shot: “A flat parcel opens into a paper city display while the camera remains locked.” Then choose the two endpoints. Only after those images form a believable transition should you create the middle state that carries the idea.

Ask what the viewer must understand before the shot ends. If the point is that the parcel transforms, show the instant when the structure is unmistakably becoming a city. Do not spend the timed anchor on a second closed-box angle.

Before paying for a request, place the three images side by side and check:

1. The same object occupies roughly the same region of the frame.
2. Perspective and camera position do not jump.
3. The middle state can plausibly connect both endpoints.
4. Only one main action needs to happen between each pair.
5. Text and logos are absent unless you are prepared to inspect and replace them later.

These checks cannot guarantee a good generation. They remove contradictions that a longer prompt cannot repair.

An original planning example

The following example is a request plan, not a generated result or a tested recommendation.

Prepare three 16:9 images from the same locked camera:

• Start: a shallow kraft-paper parcel, closed on a walnut table;
• Middle: its four sides halfway unfolded, with small paper buildings beginning to rise;
• End: a complete miniature night city standing inside the flattened parcel.

Then use a request shaped like this:

{
  "image_url": "https://example.com/parcel-closed.jpg",
  "middle_image_url": "https://example.com/parcel-unfolding.jpg",
  "middle_frame_time": 2.5,
  "end_image_url": "https://example.com/paper-city-complete.jpg",
  "duration": 5,
  "resolution": "768P",
  "prompt_expansion_mode": "disabled",
  "prompt": "A locked tabletop camera watches one kraft-paper parcel unfold continuously into the miniature paper city shown by the middle and final images. The four sides lower in a single smooth motion as folded buildings rise from inside. Preserve the walnut table, parcel material, camera, framing and warm side light. No cuts, no camera movement, no extra objects, no readable text. Audio: dry paper folds, soft cardboard taps and quiet room tone."
}

The prompt explains motion between the states. It does not repeat a catalog of everything already visible in the pictures. Keeping expansion disabled also makes the planning example easier to diagnose: if the transition fails, you know the submitted instruction was not rewritten first.

Use the correct endpoint and resolution

The three-image combination has a few non-negotiable constraints:

• it requires a first image, middle image, and last image together;
• middle_frame_time must be strictly between the first and last frames;
• the middle-frame combination is limited to native 480P or 768P;
• the prompt and prompt_expansion_mode are required by the Reference-to-Video schema;
• the returned video may run up to about 0.7 seconds longer than the requested duration.

The endpoint as a whole lists 1080P refinement, but the middle-image field explicitly limits this combination to 480P or 768P. Selecting 1080P because it appears elsewhere in the schema does not override that restriction.

The ordinary minimax/h3-max/image-to-video and minimax/h3-max-turbo/image-to-video contracts accept a first image and optional last image. They do not expose middle_image_url or middle_frame_time. A UI labelled “H3 Max” is not proof that it has wired the three-frame route.

What a five-second attempt costs on fal

On October 18, 2026, the Reference-to-Video page lists duration billing at $0.05 per second for 480P and $0.08 per second for 768P. That makes the base video charge for a five-second attempt:

480P: 5 × $0.05 = $0.25
768P: 5 × $0.08 = $0.40

Those figures exclude any billable reference-token overage. The page says one request includes 4,096 reference tokens shared by ordinary reference images, video, and audio, then charges $0.02 per additional 1,000 tokens on a prorated basis. Check the live quote before submitting because provider contracts and prices can change.

Do the first structural test at the lowest resolution that still lets you judge the important state. A sharper failure is still a failure. Move to 768P after the three pictures and timing produce a useful progression.

Diagnose the transition in the right order

When a result misses the middle state, do not immediately add adjectives.

First confirm that the request used the Reference-to-Video endpoint and passed all three image fields. Then inspect the pictures: does the subject move, scale, or change viewpoint so dramatically that the middle state cannot connect both sides? Next, adjust the middle time while keeping the images, prompt, and seed fixed. That isolates timing from art direction.

Only after those checks should you rewrite the motion sentence. A prompt can clarify “unfold continuously” or “hold the open state briefly.” It cannot make three incompatible photographs share one camera.

Review the returned length too. If a six-second deliverable must be exactly six seconds, the endpoint's possible extra tail belongs in the edit plan. Do not assume the requested duration is the final timeline length.

When two frames are enough

A middle anchor earns its place when one intermediate state carries the idea: the open flower, the profile during a turn, the half-assembled product, or the reveal before the final hold. If you only care where the shot starts and finishes, use a simpler first-and-last-frame workflow. Fewer constraints are easier to prepare and cheaper to debug.

For a sequence of separate shots, plan a storyboard-to-video workflow instead of forcing cuts into three stills. If the problem is viewpoint rather than subject state, read the camera movement prompt guide. Our H3 Max limitations page also separates the controls exposed here from features available only on other endpoints.

This site's generator currently supports H3 Max Turbo text-to-video and first-to-last-frame image jobs at 480P or 768P. It does not expose middle_image_url, so the button below creates a two-anchor draft—not a timed middle-frame generation.

The useful mental model is simple: the first image starts the sentence, the last image finishes it, and the middle image marks the phrase the viewer should remember. Use it to shape pacing, not schedule a hard cut.