MiniMax H3 Image to Video: 10 Prompts, First to Last Frame
Your still is already approved. Here's how to make it move without redrawing it: 10 copy-ready MiniMax H3 Image to Video prompts, how to write motion, and when to use a last frame.
By Eftal Yurtseven · Nov 20, 2026 · 9 min read

You already have the picture. That's the whole problem. The product shot is approved, the portrait is lit exactly right, the landscape took an hour to grade, and the moment you ask a model to make it move, it starts redrawing the thing you were happy with. Searches for minimax h3 image to video mostly come from people in that spot: they don't need a model to imagine a scene, they need one to respect a scene and add time to it.
MiniMax H3 Image to Video on each::labs is built for exactly that job. You hand it a first frame, optionally a last frame, and a prompt. It animates the still and adapts the output aspect ratio to your image. Below: ten copy-ready prompts grouped by job, how to write motion so the model doesn't wander, what to cut because the image already said it, and when another H3 endpoint fits better.
What MiniMax H3 Image to Video Actually Takes
The request is small, and every field matters. first_frame_url is required: the opening image, used as frame one, with the output aspect ratio following it. last_frame_url is optional, and it's the one that changes the tool, because passing it switches you into first-and-last-frame generation. prompt is required and can run to 7,000 characters. duration is a whole number of seconds from 4 to 15, defaulting to 5. resolution is either 768P or 2K, and 2K is the default.
Notice what's missing: there's no aspect ratio field. Your image sets the frame, so crop before you upload. A 9:16 still gives you a vertical clip, and no prompt wording will change that.
MiniMax describes H3 on its official H3 announcement as a general-purpose multimodal generation model that produces video with native stereo sound, up to 15 seconds at 2K resolution, with early testing pointing at instruction following and accurate text and brand rendering. That last part is why this endpoint suits packshots and posters: logos and labels in your first frame have a better chance of surviving the motion.

Write the Motion, Not the Picture
An image-to-video prompt has one job that a text-to-video prompt doesn't: it describes change. The model already has every pixel of the starting state. What it doesn't have is a verb. So your prompt should read like a director's note to a camera operator and an actor, not like a caption.
Split the motion into three layers and write each one in a short sentence. Camera: what the lens does. Slow push in, locked off, gentle orbit left, crane up and back. Pick one move per shot, two at most, and say them in order. Subject: what changes inside the frame. Steam rises, she turns her head toward the window, the fabric settles. Name the thing that moves and the direction it moves. Pace: how fast and how much. "Slow and continuous", "a single quick snap, then stillness", "barely perceptible drift". Pace is the layer people skip, and it's the one that decides whether a clip feels expensive or frantic.
Then match the prompt to the clock. Five seconds holds one camera move and one subject action. At 10 to 15 seconds you can sequence beats ("the camera holds, then pushes in as the light shifts"). More movement is not better movement: one confident move looks directed, six look like a screensaver.
Since MiniMax says H3 generates sound natively, it's worth closing the prompt with one line of audio direction: room tone, wind, a pour, a crowd far off. If you leave it out, you're letting the model guess.

Leave Out What the Image Already Says
This is the mistake that costs the most generations. People paste their original image prompt into the video prompt, adjectives and all: "a woman with auburn hair in a green wool coat, cinematic lighting, 85mm, shallow depth of field." The model reads that as an instruction, and instructions invite change. Now it's re-deciding the coat, the hair and the lens, and your approved frame starts to drift.
Cut anything the picture already shows: wardrobe, colors, set dressing, lighting setup, lens style. Refer to things by role instead of description. "The woman", "the bottle", "the house on the ridge" is enough, because the model can see them. Keep descriptive words only when they point at something that should change ("the warm window light slowly fades to blue dusk") or at something that must hold through a big move ("the label stays readable as the bottle rotates").
Style words go too. If the still is a watercolor, it already knows. The 7,000-character ceiling is generous, but a good prompt here is a few sentences long. There's a longer take on this principle in Image to Video Prompts: Write the Motion, Not the Picture, which applies across models.
Product Prompts: Keep the Object the Hero
Product work is where first-frame fidelity pays the bills. The rule: the camera can move a lot, the product should move a little or not at all, and the environment does the storytelling.
1. Slow orbit packshot. For a bottle, sneaker or device on a clean surface.
The camera makes a slow, smooth quarter orbit to the right around the product, keeping it centered. The product does not move. The label stays sharp and readable throughout. Soft reflections glide across the surface as the angle changes. Calm, continuous pace. Audio: quiet studio room tone.
2. Pour and condensation. For a drink shot where the liquid is the action.
Locked-off camera. Cold liquid pours into the glass from above the frame, filling it to the brim with rising bubbles. Condensation beads slowly run down the outside of the glass. The can beside it stays still. Real-time pace for the pour, then a two-second hold on the full glass. Audio: fizz, ice clinking.
3. Ingredient float. For food, skincare or anything with components.
Slow push in toward the jar. Around it, the loose ingredients lift gently off the table and drift upward in slow motion, rotating slightly, never touching the jar. The jar stays perfectly still and in focus. Dreamy, weightless pace. Audio: a soft airy whoosh.

Portrait Prompts: Small Moves, Big Life
Faces are where models overact. A blink, a breath and a slight turn read as alive; a full head swing and a laugh often read as uncanny. Write the smallest motion that sells the moment, and anchor the camera so the face isn't fighting two movements at once.
4. The living portrait.
Static camera, tight framing held. The woman breathes naturally, blinks once, and her eyes shift from the lens to something off frame left. A faint smile forms at the corner of her mouth. Loose strands of hair move slightly in a light breeze. Very slow, subtle pace. Audio: distant street ambience.
5. The reveal turn. For a profile or over-the-shoulder still.
The camera slowly pushes in. The man, seen in profile, turns his head toward the camera over three seconds and meets the lens with a calm, steady expression. His shoulders stay still. Hold on his face for the final two seconds. Audio: low room tone, a single soft exhale.
Keep both at 5 or 6 seconds. Longer portrait clips invite business you didn't ask for.
Landscape Prompts: Weather Is the Motion
Landscapes don't have actors, so the sky, water and light become the cast. These are also the stills that benefit most from 2K and a longer duration, because slow change across a wide frame is the whole pleasure.
6. Time passing over a valley.
Locked-off wide shot. Clouds move steadily across the sky from left to right, casting shadows that sweep over the hills. The light warms gradually as the sun lowers. Grass ripples in waves of wind. Nothing else moves. Slow, meditative pace across the full duration. Audio: wind, faraway birdsong.
7. The drone rise. For a coastline, city or forest still taken from a low angle.
The camera rises slowly and tilts down, as if a drone is lifting off, revealing more of the coastline as it climbs. Waves roll in and break white along the shore. Gulls cross the frame once, far away. Smooth, steady ascent with no shake. Audio: surf and wind growing louder as the camera clears the cliff.

First to Last Frame: Give It a Destination
Here's where MiniMax H3 Image to Video earns a different kind of trust. With both first_frame_url and last_frame_url set, you're not asking for motion any more. You're defining a start and an end, and the prompt only has to explain how to travel between them. That's how you hit an approved opening and an approved closing composition in the same clip.
Two rules make these work. Keep the frames compatible (same aspect ratio, similar lighting logic), because a huge jump forces the model to improvise. And describe the path, not the endpoints: it can see both frames.
8. Exploded view. First frame: the assembled product. Last frame: its parts separated in the air.
The product slowly separates into its components, each part lifting and drifting outward along its own axis in a clean, mechanical exploded view. The camera holds still. Movement is smooth and evenly paced, finishing exactly on the separated arrangement. Audio: soft mechanical clicks.
9. Day to night. First frame: a storefront or house at midday. Last frame: the same view at night with lights on.
Locked-off camera. Time passes over the full duration: shadows lengthen and rotate, the sky shifts from bright blue through orange to deep night, and the windows light up one by one near the end. Architecture stays perfectly still. Gradual, continuous pace. Audio: daytime street noise fading to evening crickets.
10. The match cut. First frame: a coffee cup seen from above. Last frame: a planet or a full moon in a similar circular composition.
The camera pushes straight down into the cup as the dark coffee surface swirls. The swirl expands to fill the frame and transforms into the surface of the moon, the circular shape staying centered the whole time. One continuous move, accelerating slightly in the middle, settling gently at the end. Audio: a low rising hum that resolves into silence.
For transitions like 9 and 10, give yourself 8 to 12 seconds. A big transformation squeezed into 4 seconds looks like a glitch, not a cut.
When to Switch to Highspeed, Lite or Reference to Video
The standard endpoint isn't the only way to animate a still with H3 on each::labs, and the schemas differ in ways that matter.
MiniMax H3 Highspeed Image to Video is the speed tier. It takes the same idea (a first frame, an optional last frame, a prompt up to 7,000 characters, 4 to 15 seconds) but outputs 768P only. The field names change too: image_url and end_image_url instead of first_frame_url and last_frame_url, so don't just swap the model string in your code. Use it to iterate on motion fast, then render the keeper on the standard endpoint at 2K.
MiniMax H3 Lite Image to Video uses the same image_url and end_image_url fields and the same duration range, and offers 480P or 768P. It's the right fit for drafts, previews and in-app thumbnails where resolution isn't the point.
MiniMax H3 Reference to Video is a different tool for a different problem. It doesn't pin your image as frame one. It takes up to nine reference images, three reference videos and three reference audio clips, plus a ratio setting, and builds a new shot that keeps identity. Switch to it when the character or product has to appear in a new pose, angle or scene rather than in the exact composition you already have. The step-by-step version lives in MiniMax H3 Reference to Video: One Character, Every Shot.
No still yet? Explore with MiniMax H3 Text to Video. Weighing H3 against its predecessor? Hailuo 2.3 vs MiniMax H3 covers that, and the image to video category lists every model that animates stills.
Frequently Asked Questions
How do I use MiniMax H3 image to video with a first and last frame?
Pass both URLs. On the standard endpoint that's first_frame_url plus last_frame_url, and the request switches to first-and-last-frame generation. Your prompt should then describe the journey between the two images, not the images themselves, and both frames should share an aspect ratio so the bridge is clean.
Can I choose the aspect ratio?
Not on Image to Video. The output adapts to your first frame, so the crop you upload is the crop you get. If you need a different ratio, recrop the still first, or move to Reference to Video, which has an explicit ratio setting from 21:9 to 9:16.
What resolution and duration should I start with?
Draft at 768P and 5 seconds while you find the motion, then render the keeper at 2K. Portraits rarely need more than 6 seconds; landscapes and transitions like 8 to 12.
Should my video prompt repeat the image prompt?
Almost never. Restating wardrobe, colors and lighting invites the model to reinterpret them. Name subjects by role, spend your words on camera, subject motion and pace, and only describe appearance when it's supposed to change.



