Nano Banana 2.1 and Hailuo 3.0 Paired Up: Video From a Single Photo

Two inputs, two models, one result. To get video from a photo, a neural network needs a starting frame and a description of the motion. A Habr write-up builds on exactly this scheme: Nano Banana 2.1 prepares the image, and Hailuo 3.0 (also known as MiniMax H3) brings it to life. The RSS announcement contains only the thesis; the prompts and examples are in the full text behind the cut. So what follows covers what the stated scheme implies and what you will have to verify yourself.

Why split the work between two models

When a model generates video directly from text, it solves three problems at once: it invents the composition, settles on the character's appearance and computes the motion. An error in the first or second task then drags down everything else. Image-to-video mode removes some of that uncertainty. The face, clothing, lighting and background are fixed before the animation starts, and the video model is responsible only for motion.

That is how the work is divided in the pipeline described. Nano Banana is known as a line of image generators and editors, and Hailuo belongs to the MiniMax family of video models. The versions 2.1 and 3.0 come from the author's headline. The article offers no independent specifications for them, so I make no comparisons with competitors. The logic alone is enough for practical purposes: you can refine and regenerate the frame several times until it is flawless, and only then launch the heavier animation step.

Motion physics: where such pipelines break

The author makes motion physics a central topic, and it is the weakest point of any video model. The typical failures are well known. Hands and fingers change shape mid-motion. Feet slide across surfaces. Fabric and hair do not react to wind or inertia. Objects appear out of nowhere or vanish beyond the edge of the frame.

This drives the logic of the motion prompt. It should not repeat what is already in the photo: the model can see the character, clothing and interior on its own. The prompt only needs to describe change: what the subject does, how the camera moves, how fast the pace is, what happens with the background. The more specific the verb and the simpler the action, the lower the chance of a physical artifact. One smooth movement is usually more reliable than a scene with five events.

The announcement gives no specific wording and no comparison of results. Look for them in the main text of the article.

What is known about the pipeline

• Input 1: a starting frame prepared in Nano Banana 2.1.
• Input 2: a text description of the motion.
• Video model: Hailuo 3.0 (MiniMax H3).
• Focus of the write-up: motion physics and prompt structure.
• Source: an article on Habr, full text behind the cut.
• Not disclosed in the announcement: clip length, resolution, generation cost, license terms.

How to reproduce the pipeline in practice

Start with the frame. Build an image with a clean composition: the character full-length or waist-up, with no hands obscured and a clear light source. The simpler the geometry, the easier it is for the video model to preserve it.

Next, write the motion prompt in short phrases in the order "subject, action, camera, pace." For example: a person slowly turns their head toward the window, the camera smoothly moves closer, the fabric sways slightly. Do not add style or appearance details to the prompt; they are already baked into the frame.

Generate several variants of the same request and compare them on concrete criteria: whether the shape of the hands holds, whether the background drifts, whether the inertia matches the stated pace. If an artifact keeps recurring, simplify the action rather than making the description more complex.

Before commercial use, check what the announcement does not mention: clip length limits, output resolution, price per second and rights to the material. These parameters determine whether the pipeline suits a production workflow or only experiments.