MiniMax H3 Image-to-Video · 5 Tests on fal.ai

Quick Summary

• Credit: Creative AI Show published the embedded test, presented by Brian. The video includes five generations he made on fal.ai and opens with a separate horse-and-water example supplied by the provider.
• The laughing portrait, the subtle background graphic, the group photo and the karate pose show visually strong results. The H3 preserves plenty of detail, produces believable microexpressions, and achieves complex body movement in select examples.
• The portal test is the most useful evidence about the model limit. The character should enter the portal, but the result brings the portal closer or materializes in front of them. The image is appealing, but the requested action was not fulfilled.
• The video does not test audio, although native stereo audio is one of the H3's main capabilities. It also does not reveal all prompts, duration, seeds, number of attempts, rejected clips, latency or cost per approved result.
• YouTube compressed playback does not allow you to confirm the 2K phrase or better. The endpoint used in fal.ai offers 2K, but configured resolution, perceived detail, and effective resolution are different measurements.
• Conclusion: this is a promising first test, not a benchmark capable of declaring the H3 better than Kling, Veo or other models. The advancement is in the combination of movement, detail and multimodal context, while adherence to actions, backgrounds and repeatability still need controlled evaluation.

Credit and method of this analysis

The embedded video was published by Creative AI Show on July 31, 2026 and is hosted by Brian, who describes himself as one half of the channel. All credit for the test examples and comments goes to Brian and the Creative AI Show.

The analysis below cross-references the transcript provided with the official MiniMax H3 page, the V2 API and MiniMax pricing table, as well as the fal.ai image-to-video endpoint page and documentation. The search was updated on August 3, 2026.

The goal is not to transform a visual impression into a scientific score. It's identifying what the clips really demonstrate, what remains to be measured and how to repeat the test with useful criteria for production.

The short answer: is it the next big leap?

Maybe it's an important leap, but this video alone doesn't prove it. The five chosen results show facial emotion, detail, atmospheric movement and body action above what many workflows accepted just a few months ago. There is enough material to justify a test in your own pipeline.

At the same time, the character does not cross the portal as requested, the karate scenario presents problems and we do not see the rejected generations. Without a success rate, we don't know whether each clip appeared on the first try or after many repetitions.

The most honest formulation is this: H3 looks competitive and deserves a place on the image-to-video shortlist. The word breakthrough should be reserved for a repeatable test of movement, identity, text, audio, cost and adherence to the prompt.

How Brian tested MiniMax H3 on fal.ai

Brian opens the MiniMax H3 in fal.ai, uploads an image, and writes a movement instruction. The video goes through five cases of its own: a portrait, a client graphic, a science fiction scene, a group image and a karate photograph.

The current fal.ai endpoint requires a splash image and a prompt. It can optionally receive a final image, inherits the aspect ratio, offers 768P or 2K and uses five seconds by default. The commercial page advertises clips between five and 15 seconds at 24 fps.

MiniMax's direct API accepts H3 between four and 15 seconds. This difference in minimum duration depends on the supplier and the interface, so confirm the endpoint that you will actually pay.

Before testing: horses are an example of the supplier

The first clip shows two horses in shallow water, with slow movement, contact between their heads, reflections and ripples. It's a beautiful demonstration of subtle movement and surface physics.

However, Brian identifies this result as the example presented in fal.ai. It was not generated by him during the test shown. Product examples tend to be selected from strong outputs and do not reveal the failure rate.

Use this clip to understand the model's visual ambition. Use the following five tests to evaluate the evidence produced by the creator.

Test 1: portrait, laughter and microexpression

Brian uses a portrait of a young Angelina Jolie and asks the face to come to life and laugh. The output preserves the overall appearance, introduces movement in the eyes and mouth, and produces expression lines that he considers realistic.

The strong point is the transition from a static photograph to an expression with small facial changes, without an obvious transformation of identity in the excerpt shown. The perceived detail also remains high.

There are two reservations. First, there is no biometric comparison or repetition with multiple seeds. Second, similarities of public figures raise questions of consent, publicity and image rights. A technically convincing result does not grant authorization to publish or commercialize it.

Test 2: animate a graph without destroying the information

The second input is a graph created for a client. The instruction calls for subtle movement in the background and preservation of information. Brian notices that the background comes to life while text and key elements appear to stay in place.

This is a valuable use case for brands because the generation doesn't need to reinvent makeup. You need to respect a blocked area. Still, the video does not OCR before and after, does not overlap frames and does not enlarge numbers or small letters.

For real work, export the first and last frames, confirm each word and number, and use a visual difference to locate drift. If the information has legal, commercial, or medical value, rebuild the text as an editable layer after generation.

Test 3: a good clip that fails the requested action

In the science fiction scene, the character should enter a portal. The H3 creates wind, moving fabric, atmosphere and a strong light presence. Brian likes the visual result.

But the semantic action changes. Instead of seeing the character go through the portal, the portal approaches, grows or appears in front of them. This is a failure to adhere to the prompt, even if the clip remains interesting.

This example teaches an essential distinction: aesthetic quality measures whether we like what happened. Adherence measures whether what was requested happened. In advertising, previs, narrative continuity and VFX, a high score in one dimension does not automatically compensate for a low score in the other.

Test 4: Turn a birthday graphic into a vivid scene

Brian starts from a birthday composite created with ChatGPT Image 2.0 from photographs of client team members. The goal is to make people look natural and present in the same moment.

In the result shown, the image gains movement without immediately losing the definition of the faces. This suggests usefulness for commemorative posts, internal communications and social pieces that start with an approved composition.

The test does not include an audit of identity, hands, person-to-person contact or continuity over multiple generations. As it involves real customer photographs, the team must confirm consent, vendor retention policy, and permission for AI processing before uploading.

Test 5: full body, kick and instability in the scenario

The last input is a photo of Brian's son in a karate pose. The prompt calls for convincing blows to the bag. The clip shows punches, kicks and quick pose changes that Brian considers to be well executed.

Body movement is a more difficult stress test than subtle facial animation. Legs, arms, balance, clothing, contact and bag reaction need to remain consistent between frames. The example indicates real progress in this area.

Brian also notices problems with background elements. This prevents a clean conclusion about total consistency. In a professional test, the scenario must be analyzed separately from the subject. And, when the input includes a minor, parental consent, data minimization and publication control are no longer optional.

What these five tests really demonstrate

The examples suggest that H3 can extract plausible movement from very different images: portrait photography, graphic layout, concept art, collage of people and sporting action. This breadth is more relevant than a single perfect cinematic shot.

They also show that intention preservation varies by task. In subtle movement, the model appears conservative. In a complex spatial action, such as entering a portal, it replaces the requested choreography with another visual solution.

The evidence supports a feasibility conclusion: it is worth prototyping. It does not support an absolute leadership conclusion because there is no comparison with equivalent inputs, blind scoring, or disclosure of failed attempts.

What video doesn't measure

Not all complete prompts, chosen duration, confirmed resolution in the file, seeds, latency, queue, number of attempts or accumulated cost are shown. We also don't see the outputs that the creator decided not to include.

Video does not measure consistency of the same character across multiple shots, continuity between scenes, OCRed text, lip sync, color fidelity, usage rights, security, or performance in large batches.

Brian mentions Kling and Google Veo, but doesn't run a side-by-side test. Observations about competitors are contextual opinion, not the result of a benchmark controlled in this video.

Biggest Missing Capability: Native Audio

MiniMax presents H3 as an audiovisual model capable of generating native stereo audio. Brian explicitly states that he is not testing sound and that he did not include audio instructions.

Therefore, this review does not inform about dialogue, voice, effects, music, environment, lip synchronization or coherence between sound and movement. It's a visual image-to-video test, not a full assessment of the H3.

A second round should use the same set of images with specific sound prompts and separately evaluate intelligibility, timing, noise, mixing, musical rights, and post-production needs.

2K configured is not the same as 2K proven

Brian describes a result as definitely 2K or better. This impression of detail is valid as a reaction, but it is not a technical measurement. YouTube re-compresses the video, and the screenshot may have scaled the file.

The documentation confirms that the fal.ai endpoint offers 2K. To confirm the output, download the original MP4 and check width, height, bitrate, frame rate and moving detail. Also compare with 768P using the same input and seed when possible.

An image can be 2K in size and still contain unstable detail, temporal blur or invented information. Resolution is just one of the dimensions of quality.

How much does it cost on fal.ai and the direct API

As of August 3, 2026, the fal.ai page for MiniMax H3 image-to-video indicates $0.26 per second at 2K. This corresponds to $1.30 for five seconds, $2.60 for ten seconds and $3.90 for 15 seconds. The input image does not add cost to this page.

MiniMax's direct table indicates $0.13 per second at 2K and $0.08 per second at 768P. In the direct API, five seconds at 2K costs $0.65, ten costs $1.30, and 15 costs $1.95, before applicable additional inputs.

fal.ai can make up the difference with integration, infrastructure, queue, and development experience. Compare the cost of the full service, not just the per-second rate. To calculate the cost per approved shot, add up all the attempts and divide only by the results that arrived at the assembly.

Realistic does not mean real, nor authorized

The tests include a celebrity, customers' faces and a child. These are precisely the cases in which visual quality and responsibility need to be evaluated together.

Obtain informed consent, confirm the vendor's terms, limit the data uploaded, and define who can save, edit, and publish the output. For a public figure, do not turn recognition into a claim of support, actual performance, or rights-free commercial use.

Identify AI-generated or altered content when context could mislead audiences. Store the input, prompt, output, model version, and authorizations as part of the production record.

How to turn this experience into a useful benchmark

Choose six inputs with different difficulties: face, full body, group, text, camera movement and physical interaction. Write one observable requirement per clip and don't change the prompt between models.

Generate four trials per condition, with the same duration and resolution. Hide the model name during evaluation. Score identity, anatomy, background, physics, adherence to the prompt, camera, audio, artifacts and possibility of use without correction.

Record cost, latency, failures, attempts and post-production time. The winner is not necessarily the prettiest clip. It is the system that delivers the most approved shots per euro and per hour within the acceptable risk.

Recommended workflow for using H3 in production

Start with a clean, large enough frame. Remove text that should remain editable, fix critical hands and objects, and choose a composition that already tells the story before the movement.

Write the prompt in blocks: subject, action, camera, environment, blocked elements, and audio. To preserve a graphic piece, explicitly state that text, logo, numbers and layout remain still and identical.

Generate a short version, review it frame by frame and only then increase the duration or resolution. Save each attempt on a simple sheet with prompt, configuration, cost, failures and decision. If the output fails the action, correct the choreography instead of just rewarding aesthetics.

Final verdict

The MiniMax H3 comes out of this test as a serious option for image-to-video. The portrait, graphic and karate action show a convincing combination of detail and movement, while the portal exposes a limit of spatial understanding that can be decisive in a narrative.

Brian is right to be excited and consider H3 for your workflow. The prudent conclusion is to experiment with your own materials, a closed budget and an evaluation grid. Five chosen clips are a signal, not a success rate.

Watch the full embedded video for visual comparisons and credit to Creative AI Show. Then use the prompt pack below to repeat the test in a more measurable way.