Quick answer: a last frame is accepted on 480P, 768P and the 2K image-to-video lane, and not on 960P, 1080P or 1440P. The model interpolates motion and framing between your two stills; it does not interpolate identity, so a costume, hair or product change between frames becomes a smear or one of the two.
A first frame alone tells a video model where to start. A first and last frame tells it where to end, which turns a generation into something closer to an in-between: the model's job becomes moving from A to B rather than inventing B.
On H3 that capability is not available on every tier, and the tiers that have it behave differently. Here is the matrix, then the practical part.
Where both frames are accepted
| Tier | First frame | Last frame |
|---|---|---|
| 480P | yes | yes |
| 768P | yes | yes |
| 960P | no | no |
| 1080P | no | no |
| 1440P | no | no |
| 2K (image to video) | yes | yes |
So the practical choice is the cheap lanes — good for storyboards and blocking — or the most expensive lane, where you presumably need the resolution and the pose in the same shot.
The middle tiers accept references and audio but not a last frame. That is worth knowing before you design around 1080P.
Two rules that come with the feature: first-and-last-frame mode needs exactly two images — the opening and the closing frame, not one image plus a description or a larger reference set — and the single first-frame mode, where one still is animated without a target, exists only on the 2K tier. On 480P and 768P, a single reference image means reference conditioning, not a first frame. First-frame and first-and-last-frame modes cannot also accept reference audio.
What the model interpolates
Motion and framing. It does not interpolate identity.
If the first frame is your actor in a grey coat and the last frame is the same actor in a red one, you have not asked for a costume change; you have asked for a cut, and you will get either a cross-dissolve-shaped smear or one of the two costumes. Costume, hair, product variant: pick one per shot.
Three things that follow:
- Keep lighting consistent between the two frames. A frame lit warm and a frame lit cool produce a visible lighting drift across the clip.
- Keep scale consistent. If the last frame is framed twice as tight, the camera move is inferred as a push-in — which may be exactly right, but it is now a camera decision you made with a still.
- Keep the gap short. Two frames two seconds apart give the model a specific job. Thirty seconds of intended action between two stills gives it a summary.
A shot-design recipe
- One action per shot. From "reaching for the cup" to "holding it". Not "reaching for the cup, drinking, and setting it down".
- Frames from the same shoot. The last frame should look like it came from the same camera, because conceptually it did.
- Describe the path, not the destination. The destination is an image. The prompt should say how the camera and the subject travel between them.
- Test at 480P. Five seconds costs five credits; a failed pose is a cheap, fast lesson.
When is first-and-last-frame the wrong tool?
The motion is an effect rather than a pose. Smoke, explosions, water, crowd — these want a prompt and a single reference, because locking a last frame locks an outcome you probably cannot draw.
The two frames are from different renders. Pixel-level mismatches in grain and lens character become a visible seam in the middle of the move.
You need continuity across three shots. The last frame of shot one becomes the first frame of shot two, but the model re-renders the intervening seconds, so the seam is only as good as your two stills agree.
FAQ
Which tiers accept a last frame? 480P, 768P and the 2K image-to-video lane. Not 960P, 1080P or 1440P.
Can I use it with a prompt and a single image? No. First-and-last-frame mode needs exactly two images: the opening and the closing frame.
Why did the costume change fail? Identity is not interpolated. One action and one costume per shot; a change between frames is a cut, not a morph.
Where this runs
The first/last-frame controls appear in the image-to-video mode of the H3 generator on the tiers that support them — the rate card at https://callirra.com/minimax-h3 lists the six tiers and their per-second rates, and the generator only surfaces the frame inputs where the lane accepts them.
I work on Callirra, so the support matrix is the enforced one rather than documentation of intent. If a tier silently ignores a last frame rather than rejecting it, that would be a bug worth reporting with the job id.