First and last frame on H3, and the three tiers that refuse it

The first paired-frame shot I ever rendered came back with a face that melted in the middle and resolved into a different person at the end.

The model was not broken. My two stills were. They were the same character, generated twice, and the two renders disagreed about her jaw — so the interpolation spent the shot doing what I asked: moving from one person to another.

Paired frames are the most precise control on the model and the easiest to misuse, because the thing they interpolate is not the thing most people think they interpolate.

Where the feature exists

Three of six tiers reject the second still. Nothing is charged when they do — validation happens before the reservation — but a shot list built on the assumption "all image-to-video takes two frames" will quietly lose half its shots.

What interpolation actually carries

Between two stills of the same subject, the model interpolates:

• Framing — a wider composition resolving into a tighter one.
• Camera position — a small orbit, a push, a rise.
• Pose — a limb moving from one place to another, within reason.
• Light and environment — a sky changing, a room darkening.

What it does not carry is identity. If the two stills are different renders of the same character, the difference becomes the subject of the shot.

The call

curl https://api.callirra.com/v1/videos \
  -H "Authorization: Bearer $CALLIRRA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "minimax-h3",
    "mode": "image_to_video",
    "prompt": "the model turns from profile to camera as the light shifts",
    "resolution": "768P",
    "duration_seconds": 5,
    "aspect_ratio": "16:9",
    "frame_images": [
      "https://your-cdn.example/first.jpg",
      "https://your-cdn.example/last.jpg"
    ]
  }'

Both images in the same array, first then last. On a tier that does not accept a pair, the request is rejected rather than silently using only the first.

The checklist that prevents the melt

1. 1.Same camera height and lens. Eye level to overhead produces a crane move you did not ask for.
2. 2.Same light direction. A key light that flips sides reads as a cut, not a move.
3. 3.Same wardrobe and props. Anything that changes will morph.
4. 4.Duration follows distance. Two seconds for a small move; longer for anything that has to read as a journey.
5. 5.Both stills from the same reference set. This is the one that matters most.

Where it goes wrong

Using 1080P for a landing shot. Accepted as a first frame only.

A last frame from a different model or a different seed. Different renders disagree exactly where the eye is: jaw, hairline, garment logo.

Expecting a cut. This is one continuous shot. It will not produce a transition between two shots.

Very long durations for small moves. The interpolation has to fill the time, and what it fills it with is drift.

When not to use it

If you need text-to-video, paired frames do not apply. If the two images describe different people, no duration fixes it. And if the delivery needs 21:9 or 4:3, this control lives on tiers whose aspect set is narrower — check the pair and the ratio separately.

FAQ

Which tiers accept both frames? 480P, 768P and the 2K image-to-video lane.

Can text-to-video take a last frame? No. Both frames are image inputs.

Will the face stay identical? Motion and framing are carried. Identity comes from the stills agreeing.

Does it cost more than a normal job? It is billed per output second like any other job at that tier.

Frame-pair uploads live in Callirra's H3 generator. If a pair works on a tier this post says refuses it, that is a documentation error worth reporting.