What Did the Reference Change? A MiniMax H3 Experiment for AI Learners

You add two reference images to an AI video request. The next clip has better-looking colours, but the room’s layout is less recognisable. Was the second image helpful? A single attractive result cannot tell you which input contributed what, or whether the change would appear again.

For someone learning generative AI, this is a more useful question than “Which prompt makes the prettiest video?” It connects a browser-based creative tool to a familiar experimental habit: define the question, arrange a comparison and keep evidence that another person can inspect.

The following exercise explores reference influence, not model rankings. It is a proposed learning activity rather than a report of tested results. Its purpose is to separate changes in spatial layout from changes in visual treatment.

Separate the question from the aesthetic preference

Imagine a fictional atrium with a circular skylight, one curved bench along the left wall and an unobstructed centre. The intended video is a slow forward camera move through that space. No people, lettering or moving furniture are needed.

Prepare two original reference images. The first is a plain monochrome perspective drawing showing the skylight and bench position. The second is an abstract texture study with terracotta and teal colours, but no recognisable room or objects. Call them the layout reference and the style reference in your notes.

Those are experimental roles, not necessarily names of controls in the generator. The drawing may influence appearance as well as arrangement. The colour study may affect lighting or introduce unwanted textures. The exercise asks what actually happens, rather than assuming the inputs will remain neatly separated.

MiniMax H3 offers a reference-to-video route with supported image inputs that can be named in the prompt. This makes the two-image exercise possible in the browser, without training a model. Keep it inside that route for the comparison; switching between reference generation and first-frame generation would introduce another difference.

Build four requests around one scene

Write one base description: “An empty atrium with a circular skylight and a curved bench along the left wall. The centre remains open. The camera moves slowly forward. No people or lettering.” Keep this scene description unchanged across the set.

For this MiniMax H3 exercise, use a short duration and the same available resolution and aspect ratio throughout. Record the visible model selection and the date. In the reference workflow, images are optional, so a text-only request can serve as the starting condition without changing the workflow.

For B, identify the uploaded drawing and ask it to guide spatial arrangement. For C, identify the texture study and ask for its palette and surface treatment. D includes both instructions. The base description stays fixed, but these role instructions change with the inputs.

That qualification matters. You are comparing reference packages, including the wording that explains them. You are not isolating the effect of image pixels independently of the accompanying text.

Decide what counts before watching

A general rating such as “cinematic” lets an attractive colour grade overshadow a missing bench. Instead, write down separate observations before generating. For layout, check the left-side bench, circular skylight and open centre. For treatment, check whether terracotta and teal are visibly present and whether the surfaces resemble the texture study.

Inspect motion separately. Does the camera appear to move forward, or does the room rotate around the viewer? Watch the entire clip rather than judging only its opening image. A layout that starts correctly but rearranges halfway through has not remained stable.

Use simple categories such as present, missing and uncertain. They are easier to explain than a precise numerical score with no agreed meaning. In a study group, let two people inspect the clips independently, then discuss where their observations differ.

Keep interpretation distinct from observation. “The bench remains on the left” reports something visible. “The model understood architecture” makes a much larger claim. The first is useful evidence for this exercise; the second is not established by it.

Repeat the comparison, not just the favourite

MiniMax H3 reference video generation gives learners a way to revisit the same input conditions. If the allowance permits, repeat every condition the same number of times. Repeating only D after it produces an attractive clip gives that condition more chances to impress.

Use neutral filenames when reviewing, and save the condition key separately. Do not discard a failed generation silently: distinguish a request that produced no usable output from a completed clip with the wrong layout. Both affect what the exercise tells you.

There is no need to invent a seed setting or automated evaluation feature. Record the controls actually exposed in the browser, the full prompt, reference filenames and each downloaded result. A notebook or ordinary spreadsheet is enough for the observation log.

Several runs still make a small exploratory sample, not a reliable benchmark. The service may change, and generation variation may remain substantial. If the outcomes overlap, say that the exercise did not reveal a clear difference instead of forcing a winner.

Finish with a narrower claim

A useful conclusion might concern this scene and these files: whether the layout drawing tended to preserve the bench position, whether the style sample changed the surfaces, or whether the combined package introduced a conflict. None of those conclusions should be written before the clips exist.

For a portfolio or tutorial, publish the question, input conditions and observation rules alongside any results you choose to discuss. Other learners can then understand the procedure even if they cannot regenerate identical footage.

The skill worth carrying into the next AI project is not a universal recipe for better references. It is the ability to ask what changed, preserve a fair comparison and describe the answer without claiming more than the evidence supports.

Sandeep Kumar

Sandeep Kumar is the Founder & CEO of Aitude, a leading AI tools, research, and tutorial platform dedicated to empowering learners, researchers, and innovators. Under his leadership, Aitude has become a go-to resource for those seeking the latest in artificial intelligence, machine learning, computer vision, and development strategies.